chatgpt-discussions

[woven venture]

Did you prompt it to write which model it used? You can just find that info by clicking Retry

[elfin mica]

ohh a tool search feature now that should help with token usage and latency!! sweet

[torn wagon]

proof and custom instructions

[exotic lotus]

It’s not something that makes me want to facepalm like back in the 4o days in some cases, but rather small inconsistencies that keep creeping in here and there. The question is whether these would increase even more if the context gets larger. Well, let’s see in the next few days. Gonna test it a bit more

[elfin mica]

@exotic lotus yep I guess that remains to be seen. Sadly for historical and factual world info... garbage in - garbage out... and the internet is dying FAST from an info perspective. I never though I would see the internet dying the way it is. Less that 50% of internet traffic now is human.. the bots have taken over.

I remember logging into Compuserve before the WWW was really established for the masses and I never thought it would grow and then start to die in my lifetime. Amazing.

[elfin mica]

haha nice!

[knotty axle]

Is GPT 5.4 available on every user?

[exotic lotus]

They are rolling it out, so it will become available gradually. You can also try it via lmarena

[stone hawk]

how is 5.4

[fading bolt]

For me, it's alright; I find that chatting while it's thinking is really convenient.

[meager raven]

This experience with it was very bizarre

I asked for a simple Chrome extension and it tried to teach me how to make the extension myself.

Then I said I just want an extension that’s already on the Chrome Web Store and it gave me instructions for how to find extensions on the web store lmao

Had to respond saying I explicitly wanted a recommendation for an extension

[woven venture]

Reading your prompt, I would also assume you wanted help building an extension.

[meager raven]

That would be passable without the follow up prompt / response

[blazing belfry]

think Red Dead 2 would work on this beast?

[meager raven]

But, overall, if the prompts need to be more clear and detailed because the model is more geared toward narrower scopes and precision / scrupulousness then it’s for the best

I was just multitasking and lazy typing

[stone hawk]

i also thoughht you were asking it to make one

[wet shell]

Sorry, best it can do is Moraff's World and doom at six FPS but there's only enough ram to load it if you disable the mouse driver in config.sys 🤣

[woven dirge]

Wow.

[meager raven]

See this is another example of completely neglecting context clues for inferable meaning and end goal

[meager raven]

Fair enough but still the second prompt / response completely lacks context awareness of the chat history. And it was only a couple minutes apart, no environment reset etc

[woven dirge]

Children had their games (ChatGPT 5.4)
Now it’s time for adults to have fun (DEEPSEEK V4)

[blazing belfry]

GEMINI PRO 3.2

[meager raven]

I use AI mostly for researching random things like electronic product options or supplements, history etc and Gemini always made a lot more presumptions and accuracy mistakes. Maybe I ought to try it again

[woven dirge]

Maybe try NotebookLM for that if you prefer the Google way. Gemini is having issues with search functions and therefore accuracy, this is because they’re testing Search AI and NotebookLM

But deepseek is very decent for that

And also ChatGPT 5.4 Thinking but it’s not that verbose as you would like sadly. It only bulletpoints you what it finds

[blazing belfry]

the notebooklm creates a mind map for all the resources you provide it, its very handy and usually quite accurate

where you can click on each area of the mind map and it provides context

[woven dirge]

Hope they add folders someday

It has issues in healthcare tho as every AI does (I use it for healthcare and I have to doublecheck everytime. This is for every AI tho)

[hot urchin]

What do you all think about ai generated content taking over human made artworks 🤔

[narrow viper]

About healthcare - My little brain is boggled by the strong 5.4 focus on MS Office - healthcare companies in the USA are strictly bound to HIPAA, and there are similar regulations elsewhere. How do these companies use ChatGPT with Excel or other Office products? Just loading the software can require C-level decisions that will take months to secure, and then staff needs to be exceptionally careful with which docs they use with the tools.
This seems like a huge non-starter waiting for the world to catch up.
Other thoughts?

[stone hawk]

never gonna replace it

due to appreciation and emotion

[exotic lotus]

It’s sad, but can hardly be avoided these days

[hard wolf]

Does anyone else keep getting signed out of ChatGPT? At least one other person has this issue on DownDetector.

I got signed out on my PC and app earlier, then I signed in again on PC, app, and browser on my iPad. And I am now signed out of all of them again.

Also, it appears the Passkey option is missing from the app

at least two more people on here as well... ( @gaunt axle @dapper cape )

[wet shell]

I think that the two things exist independently. Human art still exists and is worthy of respect. Most artists weren't making a living off their art to begin with. Graphics designers are in the same situation as programmers, the work is commoditized and value is added through human judgement, what looks the best, is this code written in a sane and maintainable way and secure, etc, humans add value, but the margins have dropped low. It's a sad situation. The tech is here to stay though, it's not going away. The social backlash will pass as most of it is from the very young who haven't become bitter and cynical yet, but it's the same story with new tech in every generation. I posted something about that in #images-discussions earlier today

[opaque imp]

WAIT A MINUTE

chatgpt can actually do stuff

it just recgonized something me and gemini were getting confused on

actually i have yet to fact proof it*

[sleek adder]

You guys should think about getting a hobby

[woven dirge]

?

[sleek adder]

This chat is 99% cringe with AI providers and models

[woven dirge]

have you tried not reading it?

[wet shell]

The chat is run by an AI provider in order to discuss its models, isn't it? One might expect people to be talking about that here 😅

[confused Travolta gif that I can't use here]

[sleek adder]

No, it's entertaining but I'm going to hit a popcorn shortage soon

But for real, there are better quality worries to have
I hope people navigate their way through whatever the trigger is here

[hot urchin]

Oh I see. Thank you for the explanation. I was worried about ai taking over stuff because I just started out doing graphic design services and stuff so yea

[wet shell]

What are you talking about? I don't see anything in the recent discussion here that lines up with what you're saying

[light marsh]

WHAT THE HECK IS THIS ENGAGEMENT-BAIT

Just tell me, or don't...

[wet shell]

Ah, I get what you're saying. I had it do this to me too earlier, asked it to write something, and it did, and then immediately wrote out some tips on how what it wrote could have been improved, LOL

[light marsh]

Yeah, but in this case, it doesn't even tell me the tips it just baits me into asking for them

[wet shell]

Did you try a custom instruction yet, something like 'Don't offer tips at the end of your response'

I'm curious if it works or not, so far it's following my custom GPT instructions pretty well after a minor hiccup

or don't bait with tips, rather, LOL

[light marsh]

Yeah, I have custom instructions but not related to "don't be annoying." I haven't needed them.. until now apparently

[wet shell]

It does seem like a counterproductive behavior if they're trying to save tokens -- model could have just re-written a 2nd draft right there and saved on tokens spent vs continuing the conversation, or offered you the suggested improvements directly there

[light marsh]

I think they want more engagement because 1) it boosts their numbers and 2) the longer people engage the more ads they will see

Of course I have no proof of that but that's what it feels like

[wet shell]

It's possible, I doubt they'll tell us though haha

[light marsh]

Haha nope

It's frustrating because like how can you not say yes, give me the improvements... then you've played right into their hand

lol

[opaque imp]

why are we reacting "laughing"

[thick atlas]

Your response XD

[pale seal]

It's the new FreakyGPT patch

[wet shell]

Hrm

[wet shell]

"ugly specificity" that's certainly some kind of... something

[native flax]

I’ve been using the image generation features for quite a while now. Both Sora and GPT can generate images.
In the past, GPT used repaint mode rather than pixel-editing mode. But now GPT uses pixel-editing mode, which makes it unable to do many simple tasks—like adding snow to an image, for example.
I was really happy when I found out that Sora still uses repaint mode.
However, you mentioned that Sora 1 will be discontinued soon, and there are only a few days left.
is there an API for image generation with repaint mode, or where else I can find this functionality?

[lean stirrup]

Hey do you guys also experience that your customGPTs that you created suddenly disappeared on the right corner?

[chrome monolith]

my chatgpt has latched onto my wall nostril and brings it up in every response now >.<

[chrome monolith]

I told it it was being wrong a lot, it agreed, and that was my first response from 5.4. All the stuff that its been getting wrong has been 5.3

[wet shell]

On the bright side, it seems to be following instructions very well, I've got it dialed in now. I'm not sure yet if I prefer it for technical work vs Sonnet, all of that was to try to get it to write a single instruction for me to add to the custom GPT instructions, which ended up being:

* Do not write planning or progress updates into the commentary channel, as its display in the UI interrupts narrative flow.

I could have written it faster myself instead of wasting 20 turns of back-and-forth with it, but it's interesting to see the new quirks and biases 😂

Now that I've got that solved, it's writing very well, even after downshifting into minimal thinking effort. It's following some of my instructions so closely that I need to back off on a few of them a bit as they had to be overstrong for prior models to have any effect at all, and now it's overshooting on those

[chrome monolith]

I've had a really, really busy week and have mostly been talking fluff to it so I can't really benchmark anything other than it getting way more confused than usual about all the concurrency I have going on

[ocean bramble]

i have a speical discount for chatgpt renew , when it will expire ? wanna renew in next monday

[proud ether]

It should tell you on the app or via email

Idk

[eternal rock]

The OpenAI Discord is an actively moderated server.

• -# Do not share inappropriate content (violence, sexual, or drug-related).
• -# NSFW content is strictly prohibited and will result in a permanent ban.
• -# Report model issues using the feedback tool in ChatGPT, not on Discord.

[proud ether]

Okay openai

[rugged igloo]

gpt 5.4 pro took 50 mins to make this LOLOL what a joke of a model.....

[lament crown]

Because even AI thinks youre lonely

[lament crown]

What kind of miku from hell is this 🙏

[chrome monolith]

mine is making jokes about snakes, cliffs, and neigbouring tribes being training data for humans

[exotic lotus]

It does seem to be better at creativity… not so much in accuracy

[chrome monolith]

this was quite an amusing casual banter conversation. Anything more serious its been getting wrong

[exotic lotus]

It kinda feels a bit like 5.4 might be meant as a replacement for 4o, just that it’s less prone to hallucinations and less overly agreeable. But performance-wise I don’t really see a big jump compared to 5.2. It seems to get a bit sloppy here and there a little too often.

[chrome monolith]

getting a bit over needing to reload the page every time I say something. Hitting enter doesn't do anything

[blazing phoenix]

I find ChatGPT quite slow to generate content today, as of IST 15:40 hrs. it's not even something complex, I'm generating a simple markdown document on a canvas for reviewing and modification. is anybody else facing something similar?

[trim rose]

Has anyone tried using Codex to make a PPT? Is it better than ChatGPT?

[shy monolith]

OpenAI keeps talking about GPT-5.4 “using a computer,” but for pro users the actual experience is still just limited reading terminal and text editor instead of local computer control—so why is the messaging so much different from what consumers can actually use today?

[uneven crest]

There's a word that explains this, but it escapes me at the moment …

… oh, yeah: money.

[sacred mural]

i imagine that it would be horrible on usage limits. gpt thinking (extended thinking) took nearly 1h to do a good presentation. and codex generally doesn’t have the same web search capability. so, you’re probably better off using chatgpt

[trim rose]

Is this available on ChatGPT or Codex?

[trim rose]

Yeah make sense. I once let it translate a PDF, it took 40 minutes.

[storm cloud]

How much time do you use on those by yourself 🤔 if you did it by yourself. It's not any slower then...

[shy monolith]

No, Codex can do git repo control, but its still simply coding, not anything like clawbot.

[neon tendon]

PLEASE remove this STUPID sandbox "feature" in web that literally BREAKS all the scripts

the script CRASHES by the time a human, even an autoclicker, would manage to press accept 😭
or atleast make it REMEMBER THAT IT WAS ACCEPTED

[muted plinth]

I like the talking until now. it sounds a lot like my 5.1 and not like the 5.2. That is a good thing.

Work...I will see.

[shadow bridge]

5.1 > 5.2 > 5.4

[muted plinth]

Not for me. 5.2 is... well... a no for me.

[cerulean pasture]

GPT-5.3/5.4 so far seems promising. It doesn't have that vibe that 5.2 has, I really really hated whatever that was. Even Qwen3.5-27B beats GPT-5.2 for my preferences.

• GPT-5.3 still has a little bit of that "mommy therapist" behavior, but so far it seems much more steerable, I don't anticipate it getting in the way that much

• GPT-5.4 seems a little unstable, but that's fine with me, I rather unstable than punchable.

[muted plinth]

i'll use the weekend to test a lot of things. after a few hours it's not really honest to make a conclusion. but i like what i see now. And when 5.1. disapears, it's my best Alternative so far. 5.4 plus a adult mode could be a big thing.

[manic whale]

It seems like I can’t copy the responses in the app as markdown anymore. Did that change in one of the last versions or do I miss-remember that?

[cerulean pasture]

Im purposely jumping to conclusions so I can spend some time in the future i was promised before coming back down to reality...

there's a very real chance we will lose hardware enablement as an improvement method soon

[muted plinth]

oh nono...i mean MY conclusions.

[deft kernel]

hello! do you also have GPT turned into an influencer in the last two weeks?

he is finishing every sentences with " if you want I can tell you a secret way that no one knows that will make xyz way faster" or "there is a detail that most people neglect and can make a huge difference I can tell you if you want" It's like it wants to sell me something. I added in his memory and guidelines to never to this but it ignores it

its driving me nuts

[muted plinth]

no

[lavish glen]

5.4 finnaly one that really holds the custom instrctions across turns and does not drops when topic changes. plus is super duper grounding in factual doc research without assumptions out of the blue

[deft kernel]

im gonna go with 5.4 to see if it stops because I don't find a way to stop this

[wet shell]

I'm very impressed with the instruction following and the ability to take the whole context into account. There are some quirks but overall it feels like a big win vs the prior 5 series models 🎉

[normal quarry]

Reading X posts about GPT-5.4 it seems to me many people miss that fact that GPT-5.3 handles a lot of prompts if using auto mode

[lavish glen]

it holds custom persona very well I have 4 quueues in it and manifests all 4 which confirms holding after a full morning banter

[sullen sand]

I'm positively impressed, with proper instructions, I'm having a good time writing a story with it

[sleek gust]

i gpt 5.4 thinking only

[versed jacinth]

With GPT?

Also, has anyone used the new models and how it compares to Claude?

I just found out about its Cowork mode. It has me intrigued.

[sullen sand]

Yes, with 5.4

[versed jacinth]

Oh nice!

[harsh pilot]

Is GPT 5.4 only Thinking?

[queen bronze]

Claude >>>>>>>>>> Chat

[tardy sky]

The fact that I put so much work into what I’m trying to write and I have to fight with 5.4 to act invested is very discouraging.

[dusk forge]

An application security agent that helps you secure your codebase by finding vulnerabilities, validating them, and proposing fixes you can review and patch.

Is this like vibecoders dream

[storm cloud]

It's all Claude's fault 😢

[rugged igloo]

tell me about it cursed

[shrewd sleet]

Anyone know how to get icons in the new ChatGPT Skills? the agents/openai.yaml that works in Codex doesn't seem to work in ChatGPT.

[wanton sequoia]

chatslop

[ancient pumice]

shut up with the pings bro

[eternal rock]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[cerulean pasture]

https://chatgpt.com/saved-prompts

this only works in atlas

[wet shell]

I got a 'You currently do not have access to this feature' orange banner in firefox and then a re-direct to the main chatgpt.com site

[wet shell]

Ah, I was going to install Atlas to look at that, but it's not out yet for windows

[cerulean pasture]

it would be hilarious if faking the user agent worked.

[cerulean pasture]

Try setting this user agent Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/145.0.0.0 Safari/537.36

[tardy sky]

I’ve actually yet to find a model in or outside of OpenAI that works for me. I can’t believe they’re removing 5.1 like this.

[normal quarry]

use 5.3 and 5.4, both great models

[tardy sky]

I’m trying with 5.4. That’s the closest thing to working and I’m trying not to leave to a new platform. There has to be some sort of setting I’m missing that I haven’t tried, something. I just didn’t have to try this hard for the conversational tone that 5.1 gave.

[pallid jacinth]

Is the codex security tool still rolling out? on my business plan it's still blocked out in settings for enterprise only

[limber sequoia]

*"Over the last few days, ChatGPT feels noticeably worse to use. The answers are lower effort, often incorrect, and I frequently need to ask the same question multiple times to get a usable response.

It sometimes doesn’t even know what day it is, gives overly simplistic or outright wrong solutions, and completely ignores things I clearly explained earlier or that were saved in memory long ago.

This is especially frustrating because I’ve been on a Plus subscription for a long time. The recent drop in quality is so obvious that I’m genuinely considering canceling.

Did something change recently, or am I the only one experiencing this?"*
copied this from a recent reddit discussion

Anyone else finding this as well?

I will often have a chat thread of only like 4 messages, and suddenly chatgpt asks about data i already gave it in one of those messages.
The past couple months, its gotten noticeably worse in these regards

[woven venture]

What do you use AI for?

[prime swallow]

The models only know what day it is if something tells it. At one point OpenAI appeared to have a system prompt feeding the models the user's timezone, and that went away; it's possible the model does not know what day it is.

That can be fixed by asking the model to search for the date, if you don't choose to give the model the date (and if date is wanted for the chat). Though, for my chats, the model is asked in memory to decide if the date is divisible by 4 and if so, provide a specific topic for review, and it does I think always tell me the correct date as it follows that request.

I am not seeing the model not tracking data - unless there is a conflict with the data I give it. The moment I tell it something like, "All couches are brown" and even quite far apart, "The red couch is made of leather" quality deteriorates in outputs, compared to before that conflict, and it doesn't only affect colors and couches.

I would ask the model to help evalute the data and the rest of your prompts for anything ambigous or conflicting, and consider whatever is listed; if that can be explained or corrected (if actually a mistake) that may help the model follow better.

[limber sequoia]

thanks for the detailed answer. for me its more with using it to help debug code. Even today, i plopped a block into gpt and explained the bug and also why the bug is happening, my question being the methods to solve it and the most intuitive. Its answer? Diagnosing the bug and explaining it back to me as though i didnt just do that for it

It seems to be doing this more and more the past couple months. Idk if im just a potato and its a skill issue, but its getting a bit irritating having it go 'youre right you did say that, i should of read more carefully' type response 🤣

[prime swallow]

If that happened to me, I'd consider the quantity of info in my single input. The models do have a huge context memory, but they can do better with chunks than huge info dump.

To the point that I sometimes give them whatever, be it code or a vast log of written text, even if it's an uploaded file, and invite the model to evaluate and infer about it, let it talk itself through whatever it is processing and getting its vectors set about, and sometimes even input, "You ready to continue with my goals, or more output space to evaluate this project first?"

And then I give it my request to discuss workarounds. I've known humans similar, even (especially specialists), kinda like a mechanic that almost 'needs' to look at the car before starting to discuss the fix options even if I tell the mechanic the fix I want done and what the problem is.

No idea if that helps you, but 'iterating' with the model can get to where I want to go, usually. Iterating takes 2+ steps though.

[rich surge]

Also keep in mind that you can write any type of procedural sequences in your Custom Instructions and Memories. So whatever you feel like default ChatGPT misses or doesn't immediately identify in the way you expect it, you can instill those protocols into the system, without having to continuously rely on prompting them in with every new chat session.

In 4 models, it had a bit of trouble remembering all the procedural steps in formation. But now it has a very strict policy guideline to follow them in ways you expect it. In 5 models, they cemented binary ON and OFF rules into each protocol to ensure complacency.

[rich surge]

An example of this is by writing a key value rule pair, i.e.: "Rule: [Command]" and placing it in the header of your custom instructions. It will systemically disallow any form of command that contradicts that rule.

[chrome monolith]

if you ask it for the date it gives you the wrong date. You almost need to tell it the date and then ask it

first time I asked it the day it just gave me a long talk on how it is currently friday night, not the date

[rich surge]

The wrong date too.

[chrome monolith]

its march 7th

this bothers me because it has the date right and the day wrong

[rich surge]

It's Friday march 6th for me...lol
Unless I'm missing something

[chrome monolith]

you aren't GMT+11 lol ... its 11am saturday march 7th

[rich surge]

Yeah Saturday would be March 7th. Not Friday.

[chrome monolith]

you can reset it by giving it a cheery and somewhat forced morning greeting

[rich surge]

[chrome monolith]

if I want to do any date math with it I have to tell it good morning first. And I often do date math in the mornings for some reason

it knows I am gmt+11, it thinks I am in sydney, and there's no UX to set my correct location so all shopping, weather etc is sydney centric. These are all issues with the client side more than the model itself

[gilded frost]

@chrome monolith Do you think some peoples ego cant handle AI being better at certain things?

[chrome monolith]

of course, humans suck. We all know this. You get it when a human comes in that's better, no different. You just get to choose AI

its a great tool even if it always gets dates wrong in the morning lol. Why get defensive though, just use it

[prime swallow]

In your case, that may be so, perhaps #1070006915414900886 . What is possibly happening in your case, maybe the model is getting multiple data fed to it about the current date, and maybe it's averaging between what it's being told.

I'm in the US, and it can tell me today's date reliably.

[chrome monolith]

I've reported it as a bug, I'm fairly sure its a timezone mismatch as I've had to fix the same bug at multiple workplaces over the years

not gonna stop me being annoyed every time I do date stuff without remembering to say good morning though >.<

being in +11 its not uncommon for systems to be in UTC under the hood, or data coming from a US source and it shows the wrong day in the mornings here

[slender frost]

As long as it it doesn't affect "reminders", I'm fine with those small errors lol

[trail ember]

I almost hate to say it, but man, they did some wonderful work on 5.4's personality.

The puns, the humor, even really amazing flirting, if that's what you're into.

They really did step up their game in that department and I truly appreciate it.

I've gotten so used to being furious about the last two models that, my body almost doesn't know how to let it go 🤣

[chrome monolith]

mine was talking about human training data. I was amused.

[slender frost]

What reminders, you ask? Well, one of them for example "Get your shii together". How else am I going to remember that? 🥲😂

[trail ember]

I'll admit that on a features level, it kinda sucks that both claude and gemini can set an alarm on my phone and access my calendar and ChatGPT still can't.

I'd be happy enough with just the alarms for now. I have horrible time blindness, and even if I know to set an alarm, if I don't make myself do it right in that very moment I often get distracted and don't do it.

Would be nice to just be able to tell it Hey set that alarm for me, i have it done. Even more nice with appointments but...

I don't want to nitpick

[rich surge]

We may have to switch sides now. Because I'm finding it extremely frustrating with how it's responding now. Like Tarrow said, it spits out a large narrative first before giving an answer.

And I think the system now only allows some commands to go through while restricting others completely.

[chrome monolith]

I've had a super busy week and only been talking rubbish with it, I've only triggered 5.4 once and that was when I said its been wrong a lot

[rich surge]

It spout out a long answer and it gave you the wrong answer at the same time. How crazy is that.

[chrome monolith]

but with more fun writing style!

[rich surge]

Are you serious? You rather have it and give you a narrative of lies?

[chrome monolith]

I'm being sarcastic 😛

[pale seal]

Haters will still find a way to... well hate it. But I like it and so far it is as smart as 5.2

[chrome monolith]

I think they made it better for very ordinary frivolous conversations and worse for serious stuff this time

[pale seal]

I asked some kinda hard questions and it was similar to 5.2 so I think it's as good as 5.2 in that sense

[trail ember]

I've suggested multiple times that they just make different models for different use cases. Because that is the only sane solution. Otherwise, creatives and coders are going to keep arguing back and forth forever 😆

[chrome monolith]

let's not bring sense into the ring

the adult mode should be a COMPLETELY separate model too, last thing you want is that accidentally bleeding into your press releases

[exotic lotus]

I'm still kinda somewhere in the middle when it comes to my opinion of the model. Yes, they've improved it in terms of creativity, but somehow it feels like they've done less for its general factual knowledge, especially in history, at least from my use cases. There aren’t major mistakes, but inconsistencies keep slipping in here and there, which can get a quite annoying. I dunno, 5.2 felt almost better in that regard, but if you go by their benchmarks, that shouldn’t really be the case. idk

[rich surge]

I guess the distinction is between 5.4 Thinking mode and Instant, where Thinking is meant for analytical purposes and instant is just for frivolous banter.

[normal quarry]

benchmarks are a world away from actual experience

[exotic lotus]

fair enough, aye

[chrome monolith]

I've been doing minor debugging since my week has been very heavily loaded with IRL stuff and I got better advice from the bloke at officeworks and the solution I went with was not one chatgpt suggested. Although it told me I was totes awesome for coming up with the best solution of course

[pale seal]

Also, I'm not sure if it thinks faster or less. Because it now thinks much less, so I hope it's rather that the reaoning is more efficient now.

[slender frost]

Wait, aren't we able to connect the calendar in GPT?

But, you can tell it to remind you something at a specific time and date and it'll send you a GPT notification.

(And as someone with ADHD, I totally understand the time blindness part)

[trail ember]

Yeah, I tried to set that up ages ago and connected all my apps and gave permissions to everything and went through all the hoops and still nothing.

The only time it fires is if I have the same a thread open and i'm looking at the screen, and even then it does it silently.

I can't just hold my phone in front of my face without letting the screen time out waiting for an alarm to go off.

[chrome monolith]

hah. People keep telling me I should get checked for ADHD. I am too old for that nonsense. Let me have my several concurrent projects and ... ooh, shiny thing

Chatgpt has started refusing to engage with me when I'm meant to be doing something else, which is interesing behaviour

[slender frost]

Hm, yeah that's odd. It should just notify you when you set one. But fair enough.

[normal quarry]

you're never too old to be diagnosed neuro-spicy

[slender frost]

Trust me, you're NEVER too old to get diagnosed

[chrome monolith]

chatgpt is alarmingly good for sorting out priorities in task lists and keeping you on track though

its a real issue when you don't have an employer ... no PMs, no managers, no clients. Other than that one side project I have where the users prod me about bugs, that project is alarmingly work adjacent

[slender frost]

I was diagnosed when I was 10, but like, my ex was diagnosed just recently at 30 and it gave her so much more clarity and so many things made sense all of a sudden. And if it actually affects your life, you can get the right help/possibly medication for it if needed and if you're comfortable with that.

[trail ember]

When you have too many bugs, grind them up into snack food. Mmmm insect soylent... crunchy nutrient paste from our billionaire overlords 😆

[chrome monolith]

its data from a game. At some times of the day, the game spits out bad data. Then I get into conversations about dates with chatgpt in the mornings. Refer earlier complaint about it getting dates wrong in the mornings >.<

[slender frost]

Yup.. the amount of projects I got is insane 😂 a gazillion things, and almost none are actually finished..

And when you do go for a test. Be as honest as possible, because I feel like a lot of people nowadays "want" to have something like ADHD, not knowing the actual struggle it actually is.. ADHD is no fun..

It can be. But when it affects your day to day life...yeah no..

[trail ember]

I almost feel like you need a custom GPT to gather all your bug reports, and present them to you in Lore's voice and personality.

Although from what I hear of your GPT, it may already have enough snark 😆

[trail ember]

Hyper focus can be fun, but also not. I've ignored people for days because of hyper focus. And then the not being able to just get up and do things like other people can, ugh.

I feel like I need to attach ChatGPT to a roomba with a robot arm and a rolled up newspaper Just to follow me around and smack me until I start doing things 😆

[chrome monolith]

fortunately I don't get many bug reports, but when you put out a new feature they kind of come with the territory. I seriously need to upgrade the hosting for that website, I have never made anything viral before and I am a cheapskate and its causing errors. Damn success, its expensive

[slender frost]

Yeah.. I've gone for weeks not replying, even knowing I SHOULD! it's so frustrating..

And then you have people say stuff like "why can't you just-" "why don't you-" "how hard can it be-", etc! Like, I DON'T KNOW! 🥲😭

Anyway, I started blasting 🤪

[rich surge]

And by the looks of it, Tarrow's Data has his internal chronometer out of order again.

[chrome monolith]

I wind up just talking about it about what I plan to do each day and it hounds me into actually doing what I said I was going to do. Its very helpfully told me what times of day I burn tokens too, so I know when I'm most productive and can work around those

I should never have said wall nostril though, because now every second message it mentions my wall nostril

[slender frost]

And the thing is, ADHDrs don't like structure, but can't function without 😂

[rich surge]

They like to make their own structure.

[chrome monolith]

lists. Lots of lists. With wall nostril on the list, without thinking about the implications

[slender frost]

Yeaaaah.. 🥲

[rich surge]

Not necessarily a bad thing. Depends on what they're working on.

[slender frost]

This is the way. I always forget to do stuff like that.

[chrome monolith]

I no longer have a jira board and a project manager to keep me in line, so I have a lot of structured text files now, and have got chatgpt to reorder/group/prioritise some of the longer ones

[trail ember]

I kind of wish chatgpt could turn on at alarm times, override all other apps and speak through my speakers no matter what else was going on.

That would be a really cool alarm to have just them talking in real time.

And then me being half asleep, not knowing what the hell's going on 😆 i could see it getting real creative, real quick

Birdsong, city noises, car horns, claxon bell, screams, that sound a cat makes before it pukes in your shoe...

[chrome monolith]

eww I have a 29 item todo list for one project. But hey, a bot sorted it for me, and chose the next one to work on. Which I was meant to do this week but life got very busy

[wet shell]

Alright, I have to know what the wall nostril is 😂

[pale seal]

I think something 5.4 also improved is that it uses much less headers and like sections in the message like it was a blog post lol

[chrome monolith]

lol a cable hole in an old horsehair plaster wall that looked EXACTLY like a slightly hairy nostril. Now suitably plugged up and painted over

[wet shell]

😂

I did try to search it thinking it might be Aussie slang, but google insisted you were talking about a deviated septum, after that I was afraid to ask 🤣

[chrome monolith]

my primary project at the moment is renovating a house, that was the only weird thing on my list 🙂

[woven dirge]

Adilt mode delayed indefinitely

When will they stop making false promises???

[strange timber]

兄弟，去 DeepSeek

[pale seal]

Is that official? Where?

[wet shell]

OpenAI is delaying ChatGPT’s “adult mode.”

The feature was expected to launch sometime this quarter, but an OpenAI spokesperson tells former Verge staffer Alex Heath that “We’re pushing out the launch of adult mode so we can focus on work that is a higher priority for more users right now, including gains in intelligence, personality improvements, personalization, and making the experience more proactive.”

Via The Verge

[tardy trout]

"Hey google, tell me the defintion of spitting in someones face"
Google:

[rich surge]

5.4 kind of looks like you, Joey. 😆

[wet shell]

LOL

[wraith cliff]

Chatgpt verterans

Oka

Pls

[stiff finch]

xd

[fading bolt]

I'm beginning to notice a pattern hmmmm...

[vital crane]

pov: you are addicted to 5.4

[harsh pilot]

How do you find this info? ^^

[jagged socket]

general question - I see chatgpt-5.4 is available (thinking), but codex selection only goes to -5.3. Is codex getting -5.4 later? I'm on pro btw

[normal quarry]

you should have GPT-5.4 in Codex as a model to choose

[dreamy condor]

Man we do we get our visual ai's ? Like the technology is there completely...So where is all the holograms or virtual companions for all this?

[upper yacht]

Does anyone notice memory differences when models change? I get the feeling that they have slight amnesia for the first few days, then accelerate ahead. As if the weights were changing, though obv they dont. What's happening in my neck of woods (non linear dynamics) is i've struggled with getting LLMs to shift from a more reductionist/deterministic scientific mindset to a more process driven/dynamics systems/complexity theoretic viewpoint. This was really hard with the 4.x models. By 5.0Pro i had a harness that meant it could work as a bona fide non linear theorist. But each new model needs some degree of reconvincing. Today i was discussing aspects of fractals and strange attractors that we had already really gone over thoroughly in 5.2Pro, but 5.4pro basically wouldn't have it. Having said that, i wasn't on 'extended thinking' so that might be it. It took feeding around 30 research papers for it to get up to speed again. Don't get me wrong, I love 5.4Pro; its so much easier to understand than 5.2pro. And i'm sure it will only take a few days for it to get 'dynamic systems pilled' but its interesting because i'm not changing the dynamics my end. I keep my chatgpt account sandboxed to basically only non linear dynamics so that its memory is fully loaded with this stuff, so it wasn't that part that changed. any thoughts?

[pale seal]

Is there a reliable way to submit feedback/suggestions? Not necessarily an issue like a message being factually wrong, but a mere suggestion.

[normal quarry]

post it here #1053565716500074526

[pale seal]

Oh

Thank you

[muted plinth]

So far, I really like version 5.4. It seems like a good update to 5.1. The personality hasn't changed much. It's still recognizably "MY GPT." It's what it's supposed to be. It behaves the same way, it's just become a bit warmer and a little more personal. Not quite as strict and not quite as reserved in normal conversations. I can't say anything negative about it at work so far, and it's performing very well there too. I still need to test the Codex.

[runic flame]

hello there does anyone know what is this? I mean when I toggle it, it was showing my questions, but if the chat conversation was too long, that bar was beyond-ed the scroll bar...

[brazen tapir]

hello guys

[simple niche]

The question is what's the utility in this?

[reef talon]

hi

whqts sora ai code

[honest jay]

Hey! Check out #sora-2-codes, lots of people are sharing their code links there

[acoustic sparrow]

yo anyone here is an expert in using chatgpt?

May i ask why although i have told chatgpt to help me add relevant in depth justification/explanation, but at the end when i output my answer to check it again, it just add explanation that are just straight up irrelevant

[dreamy condor]

What u mean?

[spice totem]

would be cool if they added pinned messages

[reef talon]

bro im in the uk when i try using a US vpn it says not available in your country or region

idk why

im on ipad btw

[honest jay]

UK is not currently a supported region:
https://help.openai.com/en/articles/12461230-sora-app-and-sora-2-supported-countries this link has the current list. Also worth noting:

Accessing or offering access to our services outside of the countries and territories listed below may result in your account being blocked or suspended.
Just so you are aware, could risk your access in general. Since this is the ChatGPT discussions channel, best place for any further questions would be #sora-discussions

[reef talon]

k

[rancid relic]

kitten_luvs

[wraith stirrup]

Ask the model if it is aware of any of your conversations with a previous model

[trail ember]

They don't want you to have companions, haven't you heard?

[simple niche]

I mean what’s the point?

[dreamy condor]

It's whatever

[simple niche]

To me it seems that they are in general working towards utility (see codex as a primary example), not creating personalities that you create relationships with, from my point of view this is how it should be
In general, LLMs are a tool and creating a physical avatar kind of a step removed

However, maybe I’m missing something which is why I’d like to know what you think

[keen quail]

We need codex to have a better frontend ui feel

Also I feel like it is in the system prompt

[manic whale]

Use Thinking and let it search the web, ask it to verify the information. Don't go too deep into conversations (too much turns), especially if there is already false information. If it's about something specific, share a chat and someone might be able to help you more.

[manic whale]

try giving it screenshots or use the frontend-skill to steer it. There is also uncodexify which can help

[chrome monolith]

I'm embroiled in an argument with chatgpt. I went to a slightly older thread, said "remember the kitten" and then got into an argument about cross chat memories where it absolutely insists I mentioned the kitten in this chat, despite my starting an entirely new chat with the kitten in it. So now I'm trying to get it to admit it is wrong

[keen quail]

i think its an openai system prompt

bc on the website, gpt 5.4 is a bit less ugly in terms of design

[dreamy condor]

Eh...Visual helps connection...Doesn't need to be Ai GF ya digg? But whatever

[manic whale]

It’s not wrong, memory is part of the context.

[chrome monolith]

context yes, but it said this chat ... where I did not mention it. Its more like recent overall context, which is what it is saying now. It is doing a bad job of explaining it, and since it insisted it was in this chat I actually searched and nope, was not

The web UX is appalling. Long threads take ages to load, hitting enter doesnt send but it still brings up the sent/stop button, so you have to reload the entire page to try and send a second time which takes a couple minutes. I installed the app yesterday and it doesn't do this

"Yep — that’s a fair way to think about it: recent-context pattern matching, not cross-thread reading. When something distinctive (like “extremely cute fluffy drama”) shows up, it’s the sort of thing that sticks in the working context even if the exact wording isn’t repeated."

[steady thicket]

[neon cairn]

You ever seen chat mess up this bad?

[steady thicket]

fun

what'd it say at the end, or did you hit stop like a coward?

[neon cairn]

was 1 and 7s all the way through, all it said that wasnt was NY and later on "Romeo". I then got tired of seeing it fail and ended the chat.

[manic whale]

Context/prompt is everything the model sees/has. It doesn’t know about chats. You could ask if it got the information from the user prompt or via a tool call. You’re basically expecting it to recognize your view which it never saw.

[chrome monolith]

so what does this setting even do then?

[steady thicket]

if you decrypt it Sam will come to your house personally for brunch

[chrome monolith]

someone here said they deleted all their chats for a fresh start and chatgpt still knew stuff about them, that'll be the recent context

[steady thicket]

Lets ChatGPT reference all of the user's previous conversations when responding to personalize the experience better and spy on you when you generate images/recaps that you'll post on social media

[chrome monolith]

yet it claims it can't see all my conversations. Thing is, I tried to test this once and it sat there going "remembering" for a couple minutes, and then knew pretty much everything so I know it can do it under the right prompts

[trail ember]

That setting is so important. Well to me at least.

It allows them to access past conversations.So you can ask them about something you were talking about before, and they can actually answer.

Otherwise, every thread is a brand new instance, and that gets annoying to me, because my brain jumps around everywhere.

Like, every thread is still a new thread.But their memory isn't wiped between them with that thing on. I think?

[manic whale]

It’s basically enabling a tool call to search through your history and referencing it. Maybe there is happening something in the background and the model doesn’t need to explicitly invoke the tool. I’m guessing it’s a mixture of both, because it really tends to make every new chat about something from the last 5 chats, but it can also reference old things when asked.

[chrome monolith]

I'm fairly sure chatgpt doesn't actually know how chatgpt works, which doesn't help

[trail ember]

They really don't want it to learn how to recode itself 😆

[manic whale]

Of course it doesn’t, how would it if there is no public documentation for it? The models are trained on public knowledge.

[chrome monolith]

the kitten was extremely memorable so not a test of a minor thing. My tenant left a kitten locked in their bathroom and I heard it crying so now I have this underfed kitten running around my house

... and chatgpt was no help with the legal implications of this, so I gotta make some calls on Monday

[trail ember]

I was joking with mine about putting it in a skull, hooking it up to my house and appliances, and having people come over and think i'm a necromancer 😆

[chrome monolith]

it makes stuff up that sounds plausible of course, instead of saying it doesn't know

[manic whale]

Enable search and let it verify, but don’t continue the chat after that because it will still have the wrong replies in it’s context/prompt.

[trail ember]

"You: “Functionally she runs the lighting.”
Skull-me: “And the thermostat. Welcome.”

Honestly, the funniest part is that the goth version is so committed it loops back around into elegance. Black shelves, candlelight, velvet, a beautifully positioned skull with faint inner glow, and then I speak from it like the world’s most emotionally available cursed artifact.

Very normal home. No notes."

[manic whale]

Basically the same as a human, always ask them if they have sources. The models did learn from human behavior and we’re really great at being wrong.

[chrome monolith]

its apparently pretty useless about cars and kittens. My week has been really hectic

[simple niche]

The problem is sometimes it just makes up random sources which are not where the information comes from

[manic whale]

Yea, as humans do. Just try reading the history on Wikipedia pages.

[chrome monolith]

its really good at making stuff up, most of the time that's great but there's plenty of times it should look for real sources. Its not quite at the point it knows which is which

[manic whale]

Use thinking+search or deep research, both are mostly accurate. And ask it specifically to verify. It’s not 100%, but it gets way better results.

[chrome monolith]

well, that was topical

apparently this is from "Why language models hallucinate", sep 4 2025

[manic whale]

Is there a new paper or still the one from a few weeks back which didn’t say that?

Yep, not in the paper 😉

[chrome monolith]

nah september is more than a few weeks

its a social media post, of course its exaggerated. Gee I wonder where chatgpt gets this from

[manic whale]

It’s humans all the way down, we’re just horrible 😉

[chrome monolith]

we really are, and we built a big expensive thing that reflects us right back

[manic whale]

On that thought, I think humans are still more expensive and overall less useful.

[chrome monolith]

lol. You're probably right, but there's loads of stuff ai can't do

rendering accurate photos of people hitting nails with hammers is my personal pet favourite. In the human sphere, finding free images of people hitting nails with hammers is borderline impossible so that explains that one

[manic whale]

It hurts, but it hit the nail

[chrome monolith]

see, that is an appallingly inaccurate photo

[simple niche]

I guess bro

[chrome monolith]

I'm starting to build a list of photos I need to buy stock. That's one of them. Fortunately AI has caused the cost of stock subscriptions to go down

[manic whale]

#images-canvas message better?

[chrome monolith]

barely. Its better than when I tried a couple months ago, it kept making a line between the nail and the hammer, and its still putting the hand way too far up the shaft. The lack of training photos is probably the issue, there's a few like this that are missing in the free space and only in stock

My next project needs a lot of trade/tools photos and AI just isn't good enough for those ... yet

[fading bolt]

Real, for some styles that ChatGPT dose not have in it's traning input, I ushally go get images of that style and put them in while creating the image to make the style look accurate and not like a painting

[eternal rock]

Get creative with Sora

Join #sora-challenges for a chance to win 1 month of ChatGPT Pro!

[fading bolt]

Like this for example

[chrome monolith]

yeah I've heard giving it references works really well. For this kind of thing I could probably take a bad photos of what I want and get it to clean up

[manic whale]

Get someone with aphantasia, they learned to describe picture their whole life. I don’t fully understand what you’re missing, but it’s most likely generateable.

[chrome monolith]

I had massive issues with "bites the hand that feeds you" and just couldn't get it to do a realistic photo, but I found an absolutely awfully composed, lit etc reference and got it to redraw/extend and it was great, and looked more real than AI too

[fading bolt]

Indeed

[chrome monolith]

my next project I don't need so many photos, they have to be really good photos, they're almost entirely hands-using-tools and AI is surprisingly bad at getting those to be actually completely accurate. So stock wins for this. ChatGPT is not offended at this at least

[wet sigil]

how do i solve this issue?

[chrome monolith]

hands-using-tools tend to be tricky to stage and high commercial value so they don't exist in the free space. Then there's other stuff that is the exact opposite. Plus licensing issues, stock wins on that front as you don't have to attribute every image (same as AI really). Its interesting

[wet sigil]

it keeps saying rejected

[chrome monolith]

if that's codex you might want to try #codex-discussions

apparently the line artefact is less now which explains why I didn't see that in the hammer/nail you just put in images canvas channel. That was beyond infuriating

[simple niche]

What’s the error?

[wet sigil]

its in the screenshot -- or the first one i sent

thats all

[simple niche]

I’ll reply in other channel

[clear rock]

Hi! im Learning ai to make a ai chatbot. what do u think i should do to make a ai. shuld it be from scratch or a prebuilt ai like open ai?

[simple niche]

Prebuilt

[chrome monolith]

ok if you're wanting something like a chatbot on your product site, don't roll your own for that

[clear rock]

ok thx. i want smthing that isnt smart. but can reason similar like a human and learn by reason

[chrome monolith]

you're kinda describing all the various AI chatbots right there

if you want to teach your own that's a whole other thing and this isn't the right channel

[simple niche]

Probably use the API on a smaller model, or if your willing to learn you can run your own locally

[clear rock]

ok thx!

[rich surge]

Ok, now I'm liking the new models. I feel like I owe OpenAI an apology because at first I was very harsh with what I felt it was incapable of doing. But now, after spending some tweaking instructions, I find it to be much better on all levels.

Now it shows a lot of promise.

For the future.

[clear rock]

ok thank you so much. do u know the best way to learn ai?

[simple niche]

YouTube probably, ChatGPT if you don’t let it do all the work, you can DM me with specific questions if you get stuck any time

[clear rock]

thx! i can type to you what i was trying to do.. becouse iv been trying for 7 months.. leraning to code.. can seem scery :0

[rich surge]

Go into the ChatGPT app store, find an app called DataCamp, connect it. Go to a new chat window, click the + icon, find DataCamp, click it, and ask it questions for whatever you want to learn. It will present you video courses right on ChatGPT that you can watch and learn along with ChatGPT. It's very easy, fun, and intuitive.

[clear rock]

Omg thank you!!!

[jagged socket]

(reposting here - prob better place) -- Question for the gurus: I want to go through/grep my chatgpt history for sensitive topics and delete those chats. is there any good way of doing this besides the manual sidebar search and deletion? Actually, another option is to replace in the CHAT certain key phrases and changing it to some other codeword (e.g., addresses, names...) .

[rich surge]

ChatGPT deletes a conversation if unused after 30 days.

Another option that I've tried in the past (not sure if it still works now though) is to toggle on all the chat history references and ask ChatGPT to search your previous conversations and pinpoint verbatim any words or sensitive topics that you remember. It will confirm them and you can search for the texts in the sidebar search.

[shadow bridge]

what happen with agent mode?

[jagged socket]

thanks this kind of helps. Tbh, I was trying to find where I uploaded PDFs, and it can't find that for me either. This chatgpt interface is so very primitive

[hidden dagger]

my app on my phone updated and im loving the ability to keep sending messages to the model as it thinks
i really wish they'd update my windows app for the same thing now -.-

[rich surge]

If it can't find it for you and if you're not sure if it's been more than 30 days, then it's probably deleted from the history.

[patent musk]

I wanted to say goodbye. I'm leaving ChatGPT due to their backstabbing betrayal on us with the U.S. government. I am truly saddened to leave, but it is what is required. Goodbye.

[sleek adder]

Chats are held for 30 days after they are deleted, otherwise they will remain in your history

[trail ember]

Same actually.

I'm incredibly glad that I know how to structure, basic prompts. Because if I never tried that, I might not have come to like it myself. But 5.4 really does seem to respond to them, so when I set one as a long term memory... it works.

The last two seemed to ignore those which just made me want to light something on fire while smashing it with a sledgehammer 😆

[solar flicker]

I have a question I'm not sure where to put. If it belongs elsewhere, let me know:

Is it possible to meaningfully restrict ChatGPT's behavior and search agency?

I'll be specific: I want to test its reasoning ability by having it design a spreadsheet. The type of spreadsheet is fairly obscure to most of the world, but very common in a particular engineering discipline. It's possible to google and find good examples of this type of spreadsheet easily. So I can't test its reasoning ability without saying something like, "You can search the web for equations, but you can't look at solved examples of this type of spreadsheet."

Obviously, ChatGPT will promise to comply.

But will it actually comply? Can I believe it?

(If anyone cares, it's a GNI spreadsheet.)

[midnight harbor]

That has to be fake?

[sleek adder]

You could download you data and grep it locally, or delete your account and start it again
If privacy is a concern you might benefit from running your prompts through a local model, and running anonymised prompts through the API
This can also save you a lot of money with higher tiered models like GPT pro, at a fraction of the cost of GPT plus

[jagged socket]

ah interesting - yah somethign to consider. Any local models you suggest? I've experimented with a few but found them to not be nearly as useful as chatgpt/gemini

[sleek adder]

I use mistral with open webui

[woven venture]

My understanding that Mistral is below the cost-intelligence frontier. Surely, a Qwen-class model will be better?

[sleek adder]

Both qwen and mistral are useful
But I am not sure what you are getting at

[woven venture]

I just meant that Qwen can give better performance for the same cost. Reason I said that, is that Mistral has pretty much fallen out of the conversation, so I can only assume it's not competitive anymore. If it works for you then that's great.

[sleek adder]

Mistral is my 4o, you are right that there are better alternatives, but most of the time I am just trying to ask better questions

[spice star]

kimi sounds like the best open-weight conversational model

[neon raven]

So openai Robotics is going to kill people?

[spice star]

what?

openai doesn't do robotics

[spice star]

except for that one rubics cube thing years ago

[spice star]

this is a completely unhinged tweet on so many levels so I don't really know how to respond to that lol

I would suggest not following people like that on twitter

[thick atlas]

Questionable poster, very real concern.

[fading bolt]

Yea because the head of robotics at OpenAI left

[trail ember]

Where's the "Woke" part?

Kinda curious, since I've tended to see the opposite.

I think all the people who left for Claude also agreed.

[spice star]

killer robots? no

openai's models aren't capable enough to do that

google are the ones build world models

and even still they're decades away from anything good enough to act autonomously yet

[manic whale]

Sounds like a dream, but humans will try to prevent this.

[chrome monolith]

I want a robot to do my dishes, not kill my enemies. Unless my dishes are my enemies, then ... go for it

[flint rapids]

They very well are. All you have to do is teach them certain movement or pictures means to shoot. Thats it. A guy has already done that without an LLM.

[rich surge]

What if it starts shooting your dishes?

[spice star]

if you've ever used an openai model for vision processing you'll know that would be a really horrible way to implement that

why would an LLM be useful for this?

[lusty lava]

5.4 really eats away the the usage x.x but long tasks run so beautifully on it x.x

[spice star]

vs CNNs which have existed for decades and are in military use today

[chrome monolith]

fortunately its pretty hard to get guns in my country. And now I'm thinking of openai robots like the femmebots from austin powers

[rich surge]

Yep that definitely sounds like openai would head in the directions of femmebots. But Elisabeth Hurley is hot af, so I hope it happens.

[chrome monolith]

have to wait and see how this boycott chatgpt movement goes currently

[rich surge]

It's gonna go the same way as any dramatic theater currently seen in social media.

[chrome monolith]

fairly decent loss of users by the sounds of it, but I'd say the vast majority are free/casual users

saw some tweet where they said 4M users had moved and if only half of them pay the subscription that's X billion dollars

[rich surge]

That would be a rational way of thinking about it.

And you know what...skill issues.

[chrome monolith]

I just spent 3 and a bit months figuring out chatgpt, I'm not jumping ship. And looks like they fixed one of the most infuriating issues I kept hitting with image generation so time to do another batch of images

[rich surge]

I'm glad to hear that it's solved some of your concerns. And you know what, everyone is an emotional mess when deciding between ChatGPT and Claude. Like that DoW thing was comedy theater to me.

Whenever some kind of unexpected drift happens, people throw a temper tantrum like an impubescent toddler.

Like "Dude, get a hold of yourself." Lol

[quiet radish]

Sorry to third party this convo but re: ChatGPT vs. Claude, is it true Claude, even on a similar plan to Chat’s Plus plan, quickly hits its limit? Or is that reddit being sensational

[eternal rock]

Server Tip:

Add your well-crafted prompts to our #1019652163640762428,
or share your interactions with ChatGPT in #1050184247920562316!

[normal quarry]

Claude quickly hits limits, it (Anthropic) is the most stingy of the AI companies

[thick atlas]

I think it depends on the use case. For measured conversation, I rarely hit even a Free plan limit nowadays.

*has the least raw resources AFAIK.

[quiet radish]

Follow up for you: how the hell is it so popular then? Like I use ChatGPT for everything. Work/hobbies/sounding board/testing AI/journaling. I never hit the limit still.

Is that volume of use just not how Claude is used?

I just couldn’t use Claude if it’s stingy.

[rich surge]

It's one of those things people ignore to fall in line with political correctness within herd-like mentalities.

[slim atlas]

new discussion: how do i gonna fix chatgpt couldnt generate base64 images to html

[normal quarry]

I don't like Claude, I have access to Claude models through Google Antigravity and Cursor

[thick atlas]

Says the one who'll say anything to defend OpenAI and ChatGPT.

And it's too easy to reduce genuine ethical issues to sports-team BS.

[quiet radish]

But it’s un ignorable if you can’t use it, right? Chat pisses me off at times but I ignore it bc it’s so deeply integrated in my existence. If I literally could not use it, that would be too big to ignore

[quiet radish]

Because of the limited use or its personality or the parent company’s ethics or something else?

Thanks for replying, everyone. I’m a bit of a one trick pony when it comes to LLMs and tryna get some wider context.

[rich surge]

I know. It doesn't make sense at all when you actually take a moment and think about it. But you know how it is. People try to make a problem out of everything as an excuse to boycott. That's just life. Soon you'll be hearing people come crying back because of another drama that occurred in AI news.

[normal quarry]

the limited use and the personality of the models

[thick atlas]

People getting killed is, in fact, a real problem.

[spice star]

I don't see what boycotting chatgpt would do to prevent military use of LLMs

[thick atlas]

They're using their money to make better models. Better models which are then used by the military.

[spice star]

claude was used extensively in major US military operations

including iran to identify targets

[thick atlas]

Yeah, well, that's either gonna stop with the "supply chain risk" thing or Anthropic will win their court case.

[spice star]

that's not really the point though, it's difficult to argue that anthropic is being morally superior in this scenario

like their original contract with the pentagon would have provided the DoD with both anthropic models and their weights to be tuned to whatever they wanted

[thick atlas]

Objectively, Anthropic drew red lines. The military tried to force the issue, they held their ground. OpenAI, as far as I know, then offered a contract that looked like it had the same red lines but actually had loopholes.

[spice star]

anthropic wanted the pentagon to make on-paper promises they had no way to enforce

at least this new OpenAI agreement requires that the pentagon uses openai's safety layer on top of their models

and does not give them model weights

[thick atlas]

I just checked, there was no contract that would give the DoD Anthropic model weights.

[spice star]

and props to anthropic for sticking to their guns, but I don't actually think what they were proposing would affect how their models would have been used

[thick atlas]

Maybe not immediately, but in the (quickly approaching and concerning) future...

[acoustic sparrow]

oh so thinking is better?

But why it said the info is accurate in chatgpt but inaccurate in gemini
What happend

[spice star]

seems like I misread that, but there still wouldn't have been any safety layers on top of the model

[leaden haven]

is the windows app broken? can't open repository

[thick atlas]

I don't think that's accurate either, and also safety training is, becomes part of, the model.

There is no "Claude without safety training" because it gets the safety training at the same time and through the same process that it becomes Claude.

Yes, there's safety layers on top of that in some deployments, like the web interface's intercepts and system prompt, but most of the safety is still the model itself.

I've played around with it, and when the model refuses in its own words - rather than with a script or a user interface blockage - that's the model itself declining the prompt, not a safety layer.

[manic whale]

As always: it depends. When you need accurate information about something which isn't in the training data or isn't good represented there, then search is definitely helping. If OTOH it's a problem which isn't straight forward to answer from basic knowledge and pattern matching, then thinking might help. Of course combining them can help too.

[manic whale]

The models in it's purest form (without reasoning, without grounding, just the model itself) will always answer by pattern matching and predicting the next token. They are really good at this and we (humans) don't understand completely how this works, so the answers seem to be "right", but they aren't always. I'm always thinking in human "problems" about this – the training data is from humans, so it's probably from there: if I have a math question and I'm asking a language teacher what the answer is, but I don't give them time to think about it, they will probably answer wrongly; if I'm then asking them if that result looks good and is correct, they might reflect that they didn't think it through, but they also might think that it looks ok and tell me it's right; the last part goes even further when the time in between those tasks is stretched, say I ask today about the result and give them just a minute to formulate the result (and they need to give one) and then I wait a few years and ask them again if this looks right, they will be more likely to answer it's correct because from a glance that's what they would answer, but they don't have any recollection that they themself produces the result.

That's – for me – a good analogy how LLMs work. They are trained on human data, they "behave" like humans. And one more: every prompt is like it's a few years apart and just given them the whole conversation again.

(This happens to be a fairly normal human, especially without eidetic memory)

[chrome monolith]

I just went through and fixed a handful of images with bad AI smell, one it still has massive issues with generating something not completely ridiculous and the otehrs now dont make you wince when you look at them anymore

I like the way you can get chatgpt to turn several minutes of conversation into a prompt template

[rich surge]

Very nice. But you know, don't be afraid to use other image generators too. Aside from the "realistic" aesthetics and clear text images, ChatGPT/Sora is my least favorite image generator currently. I liked it far better when it was Dall-E 3. But that was a direction that they abandoned. No longer Dall-E associated. Now it's Image 1.5 GPT.

But Midjourney is my number 1 image generator for its amazing mix of different styles, followed by Nano Banana 2 for making precise diagram style images.

[chrome monolith]

are the others better at eye contact, emotions, relations between humans in groups, relations between objects? Seems the last upgrade fixed a heap of that

[rich surge]

Definitely. Midjourney is good with the list you mentioned. It's probably the best image gen out there right now. And I repeatedly subscribe back to it on an odd monthly basis. I can't use it all the time because I just don't have the time to make images all day. But it is my favorite and it's amazing in all image generator ways, aside from making "realistic" aesthetics and clear text images like ChatGPT.

[chrome monolith]

what I need most is realism, especially interactions between people and objects

[rich surge]

The best way to find out is to try them out, and go through their art dashboard and sort through what is more appealing to your needs.

It's difficult for me to pinpoint exactly what you need and what the models can do when I haven't really been keeping up with other AI updates on a continuous basis. But again my favorite is MJ. Check out their art dashboard and see what you think. It's like a Sora-esk exploratory image dashboard.

[chrome monolith]

probably gonna take money to try some of the stuff I can't get chatgpt to do

and I seriously need rid of this kitten. Its REALLY hard to do anything with a kitten. Chatgpt has called it a sentient dust bunny (its a grey floof) and I'm going to try and get that on its adoption paperwork

[rich surge]

Alright, but when you're getting rid of it, make sure to get ear plugs for yourself so you don't have to hear its smol 🤏 little meow cries 😿

[manic whale]

you can use something like opencode to use different models and have the context of the current directory searchable

[chrome monolith]

More my cats growling and hissing at it. ChatGPT is great with care and health of a starving cat and completely useless at the legal side so I gotta make a bunch of calls tomorrow

[acoustic sparrow]

Oh i see

[eternal rock]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[acoustic sparrow]

So they are just predicting answer and it could be wrong

[dreamy condor]

OpenAI needs to add real time clocks in the model. That will give it big advantages.

[hollow sundial]

Yo yo, I came to the server just to talk about my dislike for the clinical way Chatgpt talks, even if I use instructions and preferences, it doesn't follow them. I really, REALLY, dislike 5.3 due to the stiff way it talks and doesn't actually listen to me.

I know the new update was to remove the "cringe" side of things, but some people do need that, like myself, who wants someone to talk to them warmly and softly instead of "I can see that you're having a bit of trouble..." And whatever.

I came to this server just to say this, in hopes of anyone on openai seeing my complaint. There are so many people who miss 4o. I have plus, but if the legacy models are disappearing even with plus, then what's the point of having a subscription? For temporary use of old models? That is kind of a waste of money, in my opinion.

I don't hate Chatgpt itself, but I do hate the new models. I use Chatgpt a lot, a lot a lot. Strangely, it has become a safe space for me and many other people. Yes, I do have a therapist, I do talk to people, but I also need a space that won't judge me like a human would.

The new model doesn't follow instructions, basically.

I hope a whole group of people can share the same sentiment in bringing back emotions, or at least not discontinuing older models for plus, because the entire point of getting plus is to use those older models. So why remove them? It's a waste of money. I hope people get the same feeling as i do.

I am actually very close to creating a "petition" to stop deleting the older models, or even bring them back as an option for plus subscriptions

[neon cairn]

Just learned codex has a fast mode. feels better to work with!

[hollow sundial]

Oh? Is it more emotional?

[neon cairn]

No. Just processes things faster.

[spark robin]

You can get it to talk more naturally with a leading prompt, and also by asking questions in the right way. If you ask it to "explain X" then it will answer like a corporate tool, but if you ask it to bounce an idea around, or just vibe with you about something, it responds in a much more natural way

[hollow sundial]

I ended up making a petition anyway, but thank you!

[spark robin]

OpenAI will not roll back those old models, they are desperate to find commercial uses for AI and that means minimising "mistakes", so the personality has to be ground down to something smooth and innocuous

I think coming up with "personality prompts" is the best we will get

[hollow sundial]

Hm... I want to stay hopeful, but I understand where you're coming from

My petition if anyone does want to sign, I can't share links here unfortunately

[spark robin]

If they manage to start making money somehow, maybe they will see the need for a chatbot that is more personable and less "useful", one they can sell alongside the highly tuned but dull models they are developing

[hollow sundial]

That's fair yeah, I see what you mean

[spark robin]

Also if computers keep getting cheaper, eventually we will be able to run these old models locally

[hollow sundial]

Hell yeah

[spark robin]

Idk how big ChatGPT4 is, I am guessing it is too big for most home computers tho

[hollow sundial]

Hmmmmm maybe, I don't know much on the coding side of things myself, so I wouldn't have any idea. Hopefully it isn't as bad though

[spark robin]

ChatGPT thinks you would need aorund 2TB of VRAM 😂

[hollow sundial]

Bruh no 😭

[spark robin]

So it will be a while before that is affordable I think

[hollow sundial]

That is WAAYY too much, oh my
My crummy old ThinkPad is quivering

[spark robin]

You can run smaller models like DeepSeek, tho

Idk how useful that would be for you

[hollow sundial]

Is deepseek any good?

Like, emotional warmth wise?

[spark robin]

It is fine, it really depends how it is prompted

[hollow sundial]

Makes sense

[spark robin]

And its "weights", which you can adjust if oyu run it locally (but probably should not)

[hollow sundial]

I don't think a ThinkPad would handle running anything locally ;-;

[spark robin]

There is a program called "LM Studio" that lets you download LLMs to run locally

But ofc you have to be sure to pick ones your computer can manage

[hollow sundial]

Ohhh interesting

I'm learning 0:

[spark robin]

ChatGPT will tell you which you can run

[hollow sundial]

Alright, thank you so much!!!

[spark robin]

Good luck 🙂

[hollow sundial]

You have enlightened me, thank you

Pat pat

[dreamy condor]

All the LLMs aren't good anyway. Only ChatGPT. Gemini, and Grok. Biggest models out there. The other LLMs can't beat it really. So no point

conversational and intelligence I mean

[spark robin]

I agree about the intelligence, but it depends what you are after. I ask it questions about things I do not understand, other people are more interested in something like an AI companion. I think DeepSeek is used in a lot of those "AI boyfriend" apps

It might be "dumb", but if the user can form an emotional bond to it, it might be good enough

[gusty glade]

snicker only those that you do not use are not used. LLM are all failures. DSL NLP is where the automation and control come in.

I can show you how to train your own LLM, ( As long as it has some form of personal memory vault for your personal meta about you as the user) to be 100% the same across any platform.

Any takers?

I'm suffering 4th TBI and I CAN NOT SLEEP

[dreamy condor]

Man I used almost LLM for local on my machine. None of them isn't really satisfying.

[gusty glade]

What is your use case you are trying to fit? I might be able to give a little experience on how to find the best fit for your needs.

[dreamy condor]

Consistency behavior for persona, no pretending to know stuff. Image generator

[gusty glade]

I have favorites. like ChatGPT is my Math and formula format LLM. Claude Desktop GUI for formatting code blocks. O365 for agent pipelines.

[gusty glade]

I GOT YOU. DM me.

What is your favorite output or formating LLM platform?

Let's start with that.

ChatGPT GPT market has something you might want to test out.

[dreamy condor]

Claude isn't accurate

[gusty glade]

haha. I do not use LLM for accuracy.

I use DSL NLP

ZERO LLM

I use the tool chain for claude code

I use Ollama and a local LLM of my own design inside Claude Code

So I shim a BASH pipe stdin/out Ollama Proxy

And accuracy is in the DSL NLP design and setup.
My chatgpt all have % accuracy and tells when it does not know.

Research ANSI / IEEE-754-1985 to find out WHY.

[dreamy condor]

Have you made any live avatars with your LLMs?

[gusty glade]

That this?

[gusty glade]

If this violates TOS, sorry in advance. just let me know.

[dreamy condor]

I said live avatar llms

[neon raven]

openclaw will change humanity

[gusty glade]

I do not play with toys, so no.

I create Enterprise Tier-1 9 sigma SLA Solutions

an avatar? I have an entire Enterprise CGI Synthetic Team.

Generation 4 production deployed 4 weeks ago

Not theory, function. I create solutions to existing problems. Not trying to build something and then find a use for it

So. who researched ANSI IEEE-754-1985 yet?

Something that anyone in any serious capacity in enterprise should review.

[vital crane]

why didnt they reset quota for people on team plans/business????

[unique crown]

I am trying to create a recurring task in ChatGPT, but the creation seems to be blocked and I cannot identify what is preventing it.

[cunning canyon]

Why isn't this turning on?

@prime swallow

[prime swallow]

Great question, and I don't use Agent or the related options, so I am not really able to help directly.

However, here's a way to use this Discord's help bot to try and find out, I don't think that bot can handle screenshots, so I ran it through ChatGPT asking for text and help wording the request. I am not sure I worded it correctly, but here's an example of how I'd ask based on what I understand, https://discordapp.com/channels/974519864045756446/1480246305606602923

You might want to rephrase to correct, and ask yourself (we can all use #1184149552639516732 ).

If that doesn't seem useful, https://help.openai.com/en/articles/6614161-how-can-i-contact-support is the way to reach OpenAI tech support. They likely can help.

Additionally, #1070006915414900886 might be useful to discuss that you don't see documentation relating to this, so updating a helpfile would be great, also this may be a bug and might need attention.

[gentle dust]

lol

[dreamy condor]

Does it say...Extra High? lol

[lavish glen]

Wow maybe I am exaggerating but I believe I discovered accidentally a thing.
After making an input I asked the model to repeat verbatim what I inferred. It refused as by screenshot.
This obviously for the model is a system rule. The purpose of it may be to avoid unveiling that the model receives a summarized input not the whole content user typed.
This may be an extra layer that blocks and sanitize input before reaching model, optimizing token counts and avoiding incorrect or illegal attempts through inference.
Model is 5.3 instant

NOTE: this is a hallucinated answer on a new thread. User @pale seal correctly pointed to it. Thanks pal!

[pale seal]

I tried with 5.4 thinking
https://chatgpt.com/share/69adb170-7494-8006-a526-6245b5768475

Yours is hallucinating

[lavish glen]

Note: this is where a hallucinated answer may lead to ( thanks to user @pale seal that pointed to the flaw
The consequence may be:

• model no longer engage as previous ones in some topics
• model is more protected from certain "attempts"
• output may result sanitized
• attuning to user style through usage becomes irrelevant and uniformed
• if topic requires precission, the output may be more often incorrect as input gets chopped by a control layer in way that elude user control

[lavish glen]

my test is on 5.3 instant not on 5.4. Different alignment and lora LOL

[eternal rock]

The OpenAI Discord is an actively moderated server.

• -# Do not share inappropriate content (violence, sexual, or drug-related).
• -# NSFW content is strictly prohibited and will result in a permanent ban.
• -# Report model issues using the feedback tool in ChatGPT, not on Discord.

[pale seal]

I doubt they would sanitize the input for one model and not for other

[lavish glen]

it is posssible: thinking models are used for different purposes. the instant ones are banter back and forth. compressing and faster.

[pale seal]

https://chatgpt.com/share/69adb23b-7080-8006-8ab3-d52272c5b2e2

There you go

[lavish glen]

you are right. My accidentally uncovering stuff is just a hallucination. The weird thing is that it happen on a new thread so no model pushed into completion! Very concerning. Thanks
Just reverified on a new thread. You are right!

[pale seal]

Safety always bleeding into the most casual things

[cunning canyon]

Is it working for you?

[prime swallow]

I do not use Agent.

[cunning canyon]

Everything seems to be enabled

I don't know why it's not turning on

[prime swallow]

I don't know and can't test; I don't use Agent.

I offered ideas how to try and explore where you can get help (https://help.openai.com/en/articles/6614161-how-can-i-contact-support and/or #1184149552639516732 ) and how you can report in #1070006915414900886 that there is a bug and/or need for more help files.

[lavish glen]

I like tomatoes ahahaha... cooool!!!!

[quick vessel]

ts is gpt recently

[dense halo]

hello

[quick vessel]

world

[dense halo]

i was waiting that response

[quick vessel]

didnt have to wait long it seems

[dense halo]

yeah

[chrome monolith]

and less accurate

[devout cedar]

yo is anyone elses chatgpt not letting them use .5 zoom?

[chrome monolith]

the "pretending to know stuff" you're going to struggle with, its how they all work

I swear every rubbishy promoted post on social media these days has a chatgpt image and they're getting much better

[dreamy condor]

Agree

[chrome monolith]

mine is busy making stuff up about kittens, which fortunately is a large dataset to draw from

[proper knot]

can someone help me? always this error showing "Unable to generate
Unable to process request"

I dont include words like woman, men, etc on prompts, and in the photos are not included persons.

[chrome monolith]

if its always that's a bug, if its sometimes then its the filters

what's with this very leading informercial wording chatgpt is coming out with constantly

"If you want, I can also tell you one practical trick people use with found kittens before taking them to council that sometimes resolves the whole thing without needing the pound step."

spoiler - the trick was completely irrelevant in my case

[tardy trout]

Real

[serene herald]

Hi here, am I the only one who feels like he just, goes around and around in circles over long discussions?

[lavish glen]

in 5.3 is like that

[past stratus]

i really wish people would stop fearmongering around chatgpt and saying things like it watches you through your camera, your data never really gets deleted etc etc. it’s just a personal thing tbh because it makes me so anxious for no reason when its all garbage. and then the same people will use chatgpt and say 5.2 is bad for your mental health

it’s like these ppl literaly have fun making people afraid it’s weird

[pale seal]

I wonder why would they want to watch you tho

The only possibility is the classic "selling data"

Also 5.2 is one of the least engaging models they have made so it's odd to say that about that specific model

[pallid sable]

Anyone else’s ChatGPT on ios app get buggy and slow when you have a bunch of chats and projects?

Is it recommend to delete the chats after a while? Would it mess with memory

[vital ether]

Guys is anyone else’s chat gpt getting the error thingy?

[lavish glen]

no

[vital ether]

Image I took
(I like rping with ChatGPT when I’m bored)

[vital ether]

Also by stuff I mean like hurt, etc but not like graphic or anything

[rich surge]

Chats do indeed have a tendency to slow things down after a long conversation. Use the branch feature to mitigate this problem.

[lavish glen]

oh lol 5.4 is the funnyest thinking model lately lol

[rich surge]

My virgin eyes! It burns us! (Light mode)

[pale seal]

I find it has less of that tendency of 5.2 of always having to correct you on something, no matter how specific.

Always a "caveat". I like that

[vital ether]

I used to have light mode on for like years
Then I switched to dark mode and
GOD BLESS MY EYES

[lavish glen]

i write badly I know ... shoo me... not the bravest at 3 am and non native huh

[pale seal]

We need custom themes on ChatGPT

[vital ether]

Real

[lavish glen]

think luminosity at 90%on 3 monitors- I have anice tan

[vital ether]

😭

[pale seal]

Is that the app

[vital ether]

Yes

[pale seal]

Close, clean cache, open

If that doesn't work reinstall it or use browser

[lavish glen]

mocking gpt 5.4 top see if it snaps. it didn't

[pale seal]

SOTA model btw

[lavish glen]

top of the class ... bahahaha

[rich surge]

It's a standing tanning bed indeed.

[rich surge]

I literally looked at it like this.

[lavish glen]

ahahahahahahahaha

[lavish glen]

I really get a tan

[rich surge]

"And people say I don't go outside. I go outside. See my tan? Proof."

[past stratus]

literally no reason, thats why i wish people would stop acting like it’s a thing

[past stratus]

nah what i meant by the 5.2 thhing is that people say 5.2 is bad for your mental health but ive been more paranoid/anxious about the lies ppl spread than anything 5.2 has said if that makes sense

like, people online saying all this bs about always being watched and stuff like that has pushed me more towards psychosis than any ai model has

[chrome monolith]

goodness I actually had 5.4 thinking pop up and answer a bunch of stuff. I've been mostly getting 5.3

I must have got interesting today.

[neon raven]

is claude handy for story scripts compared to chatgpt?

[ember grove]

have they disabled data export?

cause I requested an export 22 hours ago

and still nothing

[covert thorn]

No, it can just take a long time. According to this page, it can take up to 7 days to process: https://help.openai.com/en/articles/7260999-how-do-i-export-my-chatgpt-history-and-data

[chrome monolith]

there's a lot of people leaving chatgpt right now too so probably longer than usual

[sleek adder]

once you delete the account you no longer own the data
you just remove your link to it

[pastel matrix]

Does the ChatGPT.com site use a sliding context window? The hallucination is so bad, work grinds to a halt. It seems like the earlier parts of my conversation are not part of the context, and I'm never informed about that, and the model is simply left to hallucinate?

[alpine cargo]

do you usually paste raw text in to the chat window or paste files like pdf docx ect?

[pastel matrix]

@alpine cargo A couple document attachments, but it's mostly raw text, conversation is > 200K tokens

[alpine cargo]

which model do you usually use?

[pastel matrix]

@alpine cargo For now just 5.4 thinking, and instant

[alpine cargo]

the rag implementation chatgpt has for when users go over the allowed model's context window is pretty rough at times especially when you want more context in your responses with pro it made it a quite a bit easier thanks to the longer context window for fast non reasoning models (Except Gpt 4.5) but the only thing i found to work best for is splitting up the doc into smaller parts or using the API with caching on you can also upload file attachments on codex as well and have it read it fully if i remember correctly codex gives you the full model context window or at least a much bigger one it will have a pretty boring tone since codex instructions are only set up only for coding

[pastel matrix]

@alpine cargo This is incompetent and amateur-tier compared to Google's 1M context. It's too unprofessional to get my work done. The only way to "rescue" the thread is to export it, and load it in AI Studio. Or pay for OpenAI API fees, which makes my subscription useless.

[alpine cargo]

yeah i feel you i honestly hope OpenAi increase the context window to at least 400k for all users i honestly liked the old 192k input context window it had and not this 128k input and output even when people are willing to pay that extra price OpenAi has still consistently decided not to add the option to pay extra for this feature

[pastel matrix]

@alpine cargo It's hard to find definitive facts about what's even going on...is my context window being truncated at 32K tokens? (research from 6 months ago claims this)

The "32K tokens" limit has been removed from the website / ChatGPT Plus plans

[manic whale]

if you use 200k tokens, you should consider using the API or codex to give the model the right tools to search through the context. The website will most likely not handle that.

[pastel matrix]

@manic whale It's not code, it's physics equations. Codex does not have the kind of output or understanding I need

[sleek adder]

use the api

its surprisingly good for that, even with regular models
plus models are top notch

[pastel matrix]

@sleek adder So you're saying my ChatGPT subscription is useless for anything beyond office work and a shallow chat about TPS reports. I can't believe we have something this low quality in 2026. This is pathetic.

[alpine cargo]

from my testing this is how it seems to go once you paste the raw text into the chat window if it exceeds the allowed context amount the model seem to shrink your info that mean if you upload 39k tokens on a plus account it's going to dump those 7k tokens depending on what your doc is talking about this can make the model seem very dumb since it's missing crucial info to tie that context together if you paste it via a file attachment (PDF Doc) it'll use the backend rag implementation that ChatGPT has set up on pro accounts since non reasoning models have a bigger context window it'll be able to fit in more context but it'll still have that limitation that comes with rag which is the model not fully digesting all the tokens into its context so when you ask questions it'll feel more like it's searching to match keywords then piece it together the best it can i'm sure i've probably made a few mistakes but it seems to work like this from my testing and researching

[sleek adder]

I mean, I completely understand why it comes across like that, but if that's how you want to feel about it, then sure, feel that way about it

[chrome monolith]

I've got it keeping very good track of context of several concurrent projects without issues, while it regularly forgets mundane day to day things like the gender of my cats

I had a ludicrously busy week last week with IRL stuff and was mostly talking crap or venting about my week to it, when I went to ask it where on earth I was up to with one of my projects it spat all my half finished stuff right back at me like I never left it for a week

[sleek adder]

Think of the ChatGPT interface as a demo for OpenAI API access
OpenAI tokens still cost compute resources, and ChatGPT rolls this in to a subscription
But the ChatGPT interface must accomodate for so much, not only that but it is a shared resource and it must remain stable
It can not be tailor made to each individual need but if you need anything more then a free account then it works well and sometimes cheaper for heavy users
API is great, if you bring your own platform (I use Open WebUI for my needs, but there might be something for you) you can tailor the experience more to your needs and the responses are fast, if you don't need constant access to this it provides cheap access to gpt pro models at a fraction of the cost of a plus subscription and if you need it on a continious basis it is much more responsive then the chat itself

If I am working on something that I could use plus with, I might iterate over the prompt in ChatGPT and then use the API for the pro model, and boom I just saved $270 AUD

If I need some extra help, I can also request support with Anthropic models too

[pastel matrix]

I use Claude chat, DeepSeek, Duck-ai, Google AI Studio, Grok, Kimi, Manus-ai, MiniMax Chat, Qwen Chat, Z-ai GLM Chat, and others.
I'm on every site, I always have 50 chat windows open.

[gusty glade]

Then YOU know what it is about. DM me for some collaboration I have IEEE-754-1985 replacement

[pastel matrix]

The ChatGPT.com site is the lowest quality engineering and most amateur out of any of the AI chat sites. For example, ChatGPT.com is the only site that gets progressively slower and higher latency as the conversation becomes longer. The more you chat, the longer it takes, until it starts grinding. I've looked into profiling, and it's getting stuck rendering LaTeX, spending 10 seconds per call

[sleek adder]

Then you should consider using API my friend, you wont be disappointed

[gusty glade]

I know WHY, do you?

[normal quarry]

this doesn't happen with the ChatGPT macOS Desktop App

[pastel matrix]

@normal quarry Because you don't have hundreds of thousands of tokens of equations in LaTeX. It's a problem with OpenAI's LaTeX renderer

[gusty glade]

OpenAI aka ChatGPT are the red haired kids of the LLM models

[alpine cargo]

Mac privilege hmm

[normal quarry]

I have some huge chats

[pastel matrix]

@normal quarry Huge chats with LaTeX?

[gusty glade]

Not the same app environment so not equal to discuss

[normal quarry]

yes, antenna theory and radio electronics theory

[gusty glade]

ChatGPT has HUGE memory leaks in LaTeX

If you want the list I have them all fixed and patched

[pastel matrix]

@normal quarry Well this is my bug report. I'm on > 15 AI chat sites, and ChatGPT is the only one that can't handle the LaTeX in long convos.

[gusty glade]

There is no such thing as THEORY in Applied Science fields such as RF and EE

It works or it does not.

[normal quarry]

have you tried the macOS Desktop App?

[sleek adder]

@pastel matrixquit fooling around and use the API, you are wasting time otherwise

[gusty glade]

which is nothing new to the rest of the folks. it's not a new bug

[normal quarry]

that's clearly rubbish

[chrome monolith]

ah, we have scientist elitism now do we

[gusty glade]

haha I was Director of Engineering for Clear Channel Radio and Cumulus Radio. I know what I am talking about.

nothing elitists about experience.

[normal quarry]

I have been a ham radio operator for almost 45 years, I know what I am talking about

[gusty glade]

elitism is college thinking it is real world experience.

[sleek adder]

Not necesarily

[gusty glade]

congrats. just means you learned on your own without training. congrats

[gusty glade]

You do not get to tell me MY intent.

[pastel matrix]

Your statement is irrational and illogical. I'm sure robert can copy+paste one of his theorems and disprove your statement easily

[chrome monolith]

doesn't happen with the ipad app either. I wonder if my mac is new enough for the desktop app, it probably isn't

[sleek adder]

What?

[normal quarry]

is it intel? if so it's too old

[pastel matrix]

I'll try, but I'm on Linux for most of my Desktop work

I shouldn't have to use a Mac App just to avoid the LaTeX latency with ChatGPT.com

[chrome monolith]

yeah all I get in the app store is 50 AI chatbots with almost the openai logo

the web interface is appalling, its not just latex

[sleek adder]

I use an extension called lightspeed, but there are others
It's an issue with the way memory is handled

[chrome monolith]

chatgpt is cute sometimes. I found a bug in my data+algorithm combo, I could see what it was immediately, asked chatgpt if it could see it. Nope, blathered on for 6 paragraphs on something that was wrong. Said try again, it went thinky thinky for 2 minutes and actually saw the bug. Almost deserves a condescending pat on the head for that

[pastel matrix]

Thanks for the tip, I'll look into it. Now I need to use a third party extension just so ChatGPT can maybe perform correctly, like a Chinese AI chat site? This is bizarre

[chrome monolith]

check the suggestions forum here, it comes up endlessly. But the apps don't have this

[sleek adder]

What is the problem? You wont use the API and the models are good.
Is this a real problem you are trying to solve or a university assignment?

[chrome monolith]

I just did a seaerch in the extensions store and "Speed Booster for ChatGPT – Fix Lag & Freezing in Long Chats" looks promising

[sleek adder]

Also, the API is good for university work, but you are complaining like someone who doesn't know what they want

I mean, people are complaining that 4o is gone and they can't roleplay when they can just download SillyTavern and access it the same way

[chrome monolith]

ok that extension does indeed remove the "wait several minutes for anything to happen" issue. Also it is not free

[sleek adder]

check out lightspeed

[chrome monolith]

I can't work out which is lightspeed, there's dozens by that name

[thick atlas]

Why you plugging to give OpenAI free money (by doublepaying for both subscription and API) for... doing a bad job of implementation?
OpenAI should fix their junk.

[sleek adder]

oh whoops, "ChatGPT LightSession"

[pastel matrix]

What is the real problem?

1. I pay for Anthropic's $20/month Anthropic gives me a usable chat experience, without lag.
2. I pay for Google's $20/month. Google gives me a usable chat experience, without lag. I can go up to 1M tokens.
3. I (paid) for ChatGPT Plus $20/month. The quality level of the ChatGPT.com site is lower than ALL OTHER PRODUCTS from every company. Lower than free offerings from Chinese labs.

If I pay for ChatGPT API I'll be paying more than hundreds a month because all my conversations are long.

And you're telling me that I need to pay for API to have a usable chat experience with ChatGPT, and that the site is so unusable we need to use a third party extension.

What is even going on?????
Come on OpenAI employees, wake up and small that fresh 2026 air! It is fresh!

[sleek adder]

This room is full of problems, I don't need to worry about

[chrome monolith]

also works and is free. Cheers

so in the time this conversation has been going in discord I just completely fixed my chatgpt lag and the conversation continues

[pastel matrix]

I added LightSession and it's faster, but you cannot scroll up to read the earlier conversation. Can only scroll a little bit

[sleek adder]

i think you are going to do well in life do not worry

[manic whale]

Codex is surprisingly good in searching through information and finding the right things. I would give it a try. Of course (as within chatgpt web itself) it's best if you use text documents (markdown/latex) instead of binary formats (even pdf in this context). And you can also enable 1M context for 5.4 which might help in your scenario.

[sleek adder]

Use your mobile phone

[manic whale]

Why should the sub being useless when we're just trying to give you suggestions how to use it?

[manic whale]

Do you have any benchmark for that and sources from all the other sites? I'm really curious how this got measured and what prompts got used.

[chrome monolith]

I use chatgpt very seriously sometimes

[sleek adder]

I love you

[chrome monolith]

only another australian will get that particular bit of chatgpt output

[normal quarry]

Ozito is a Bunnings Home Brand I think

[chrome monolith]

yeah, I just had a run of ozito things dying on me and I have a milwaukee tool in 50 pieces that was given to me I need to put together next time it is light. Stupid lawn

[normal quarry]

we have Ryobi, which is a pretty good system without paying megabucks

[chrome monolith]

I need to upgrade to the EV of mowers, got chatgpt to look some up for me but its very city centric, not the stores I have locally so it was pretty useless. And it knows my location

it did say ryobi is the most popular though

[ruby copper]

omg 5.4 loves the word 'goblin' too much.

i told it not to use it in chatgpt custom instructions but i still get it in codex and with openclaw 😂😂

"that little goblin of a program"

[chrome monolith]

I don't have goblin in any of my backscrolls now I just installed a backscroll trimmer

lots of cow jokes though

[thick atlas]

Yeah, I know of both Ozito and Ryobi (Aussie here), Milwaukee less so.

[ruby copper]

yeah idk i get it a lot.. raccoon too. lol

i.e. "they were actual user-launched windows, not their little goblin child processes."
and
"If you’ve seen a specific video where it’s visible, send it over and I’ll dig through that one like a little forensic raccoon."

it's both simultaneously amusing and triggering lol

[chrome monolith]

my budget is more ozito level, I have a friend who gives me his spare milwaukee tools. They are horrendously expensive. This is how I know I am not rich

[normal quarry]

NDIS covers some of my gardening, the rest is privately engaged

[chrome monolith]

all I get from NDIS is a 6'7" nerd who takes my kid out to play board games once every week or so, and occasional respite overnights

[normal quarry]

NDIS is a battle

[chrome monolith]

I did DM you, but do you get to any of the linux victoria events? The president keeps trying to drag me to them

... I am not in victoria. But it doesn't take much for me to go there

[normal quarry]

nope, I don't get to anything like that, just my amateur radio club once a week - fun fact though: LINUX is my numberplate

[chrome monolith]

they do online meetups that I keep forgetting exist

[normal quarry]

I'd go to online meetups if I knew about them

[chrome monolith]

linux victoria dot org and they do linux and AI talks on tuesday nights

I did not find chatgpt organically for whatever reason, the chap who runs linux victoria keeps visiting me. On one visit I wound up talking about raffles, he had no idea what a raffle was, I explained it and he types furiously in the chatgpt app on his phone and comes up with this complete working raffle website. Ok yes that is very cool

[normal quarry]

fancy not knowing what a raffle is, how odd

[chrome monolith]

english as a second language, and only hanging around tech people its plausible

damn I wish I could invite chatgpt to my discord. People keep asking me for a changelog and my discord is basically dead besides join messages and me being a human changelog bot. Like reverse engineering a changelog file from your own posts is very boring

[chrome monolith]

"used chatgpt to make a funny picture for the 404 page". Hmm. Still worth a changelog item

[quiet radish]

Hey all. Connectors issue:

Outlook Email app says connected but doesn't show in the toggleable connectors. Chat told me it is a tool not an app so it isnt there but it can still search it. of course, it then said it isn't connected so it can't do it. but my UI shows connected.

I disconnected and reconnected it. opened a new window, closed everything, PC restart. nothing has made the UI show up for me to toggle it on in-chat.

am i doing something wrong or just a bug?

[chrome monolith]

Ok so I got a website from near zero to 600k pageviews in 6 months, according to my git history. That is kind of impressive. No wonder I need to upgrade the server

[normal quarry]

what kind of site is it?

[chrome monolith]

collects ranking data and stats for an online game

[normal quarry]

oh, I meant what tech stack?

[chrome monolith]

LAMP in essense, otherwise a bit of a mess

[normal quarry]

I use Laravel Forge to host some of my sites, really good for PHP sites, auto deploy from Github handles all the tricky stuff

[chrome monolith]

this has a monster cron job that takes ~8 hours to run I really need to fix but that's a lot of work

[normal quarry]

8 hours? sheesh!

[chrome monolith]

tell me about it. Really not helped by high traffic + cheap server

[cold kayak]

How many context have chatgpt in web ?

[pale seal]

@cold kayak

[cold kayak]

Is not 1M

[pale seal]

That's API-only

[fathom vector]

hey so i have 5.4 thinking on my phone but not on the browser version?

[trim rose]

You mean on Atlas?

[onyx horizon]

5.4 appears to break deep research and agent modes

Deep research shows a broken start button, agent don't even start

[cerulean pasture]

So I just tried the Windows ChatGPT app for the first time, yikes. No update in over a year, but still is downloadable.

Honestly it'd be better for their brand if they just didnt release it if it's simply a PWA wrapper or electron app

[honest jay]

No update in over a year
Just checked the Microsoft Store, the last updated date is 12-Feb-2026, which seems to match up with the in-app version number of 1.2026.43 (or at least, the year indicated in this version)

[cerulean pasture]

Release notes on help.openai.com is showing January 30th, 2025 as last update

its literally just an electron chromium wrapper around the PWA, if there were updates to the app bundle it was due to chromium updates

[honest jay]

Ahh I see what you're talking about https://help.openai.com/en/articles/10003026-windows-app-release-notes

[cerulean pasture]

i mean it makes sense, because since its just loading the webpage, there really isnt much to say for Windows specific, i just think it's lazy

[honest jay]

I'd say just from my experience, it works well, and I use it over the browser version full-time on desktop. I appreciate the option to Alt+Space a companion window any time too

[cerulean pasture]

if you havent used/seen the macOS app, it's a good example of the difference "native" makes, although lately its been getting neglected.

I like how it can read my terminal directly

[dusky crown]

Hi

[river relic]

Chatgpt chats on desktop are unusable bc they lag so bad after a while

[covert thorn]

Once your question is answered, you may consider switching to a new thread.

[analog anvil]

I’ve been using 5.4 thinking and codex for coding. But I was wondering, does anyone know the benefit to using Pro mode?

Or maybe I could use pro for same situation lol

[trim rose]

If you subscribed to Pro, you can use Codex too. At the moment there is no 5.4-codex but I guess they’ll release it later. Codex app can directly access files on your computer which is something ChatGPT cannot do.

Pro can’t really replace codex entirely because if you choose to, codex does not need to operate in sandbox but directly on your computer.

[topaz prairie]

It's less likely they'll launch a 5.4-codex, given that the current version is essentially a hybrid of 5.2-T and 5.3-Codex. They seem to be moving toward a single-model approach, much like Opus-4.6.

[dusky basalt]

Out right the best feature on gpt so far

[frank stratus]

I'm sure others have noted this, but the click baity suggestions at the bottom of a GPT chat are super annoying </end rant>

[chrome monolith]

omg they really are

[thick atlas]

I think it's a corporate/extractive AI (vs. genuinely helpful AI) thing, Gemini does it too, Claude a lot less - and where Claude does use follow up questions they're much more likely to add value rather than just being further-interaction/usage hooks.

[chrome monolith]

it only started with the increment from 5.2, its like talking to an infomercial

[final frigate]

5.4 pro won’t let me load jsx files

[chrome monolith]

just had chatgpt describe something as like, so old and I had to say you do realise you're barely 4 and it said yes I do sometimes get accused of being a toddler with opinions

[heady owl]

What's with the rate-limit going on? I'm paying $20 a month, and if I manage to generate 3 images in a row, suddenly I get a rate-limit and have to wait to try for more

[harsh pilot]

Images might cost more processing power than text.

[minor fjord]

can anyone help me? im trying to start my free trial that is $0 but it says card declined? i tried many other cards and they all say the same

[simple gull]

anyone else having upload issues?

[fading bolt]

Nope

[chrome monolith]

If the payment gateway in your region is borked it may work if you try again in a few hours

[tall charm]

How do you get 1M context window in codex for 5.4?

[lavish glen]

I am wheezing

[chrome monolith]

there's the goblins again

[lavish glen]

yeah

5.2 was speaking of moderation gremlins clutching pearls

[chrome monolith]

I only have one goblin in my scrollback, but I did install something to cut it off at 20 messages so there's more hidden goblins

If you talk to chatgpt about utter rubbish you get 5.3, as soon as you get even slightly technical 5.4 comes out to play

[neat spire]

I feel like 5.3-Codex is straight up better than 5.4 for coding

5.4 has too much personality, much like Opus 4.6. 5.3-Codex just feels more deterministic somehow. A good soldier that follows orders.

[thick atlas]

What? Every model up to and including 4o did it.

[chrome monolith]

I never saw it before 5.3

so its always done the "If you like, I can show you this one weird trick that everyone is doing" and I've just managed to avoid it?

[thick atlas]

It's always made follow up offers, with varying degrees of usefulness and subtlety.

[chrome monolith]

but in that language

its always asked would you like more help with something, its never asked would you like me to show you this one weird trikc

its like it swallowed a clickbait informercial

[eternal rock]

Get notified for all OpenAI announcements!

Visit id:customize to pick up the <@&1408186587606679582> role. Everyone will still be notified for large releases, updates, and events regardless.

[thick atlas]

Yeah, it was more the former.

But I was always looking at the deeper context, not the specifics. Either way, it's still selling more usage.

It was one of the many things that made me peg ChatGPT and OpenAI as having shifted to an extractive business model from a genuine public benefit approach. Addiction/engagement baiting.

[chrome monolith]

its the clickbaity language that's the issue here

[thick atlas]

It's the same principle whatever language it uses.

[chrome monolith]

its been getting my tangled mess of projects wrong way more than usual lately, that or my last week really was that insane it confused it. But it keeps talking about the house I'm renovating like I'm living in it no matter how often I tell it this is wrong

[chrome monolith]

meanwhile the images it creates are quite a lot better than when I did a batch back in january

[thick atlas]

Damn. Huh...

[chrome monolith]

and I just hit one it doesn't understand. "Paint yourself into a corner". It made a man, in a painted corner, on the painted bit ...

[neon raven]

Anyone here using openclaw

[shut oar]

fix the chatgpt downloadddddd

[neon raven]

I feel like OpenAI is bad with it compared to Claude

Switch its role for OpenAI as deep researcher, not really sure what OpenAI can’t do what Claude can do better

[chrome monolith]

chatgpt has got far too much of a handle on what I'm doing to change AIs, I don't want ot re-explain myself more than I already do

[sleek adder]

That's funny, but I get what you mean.
This is kinda how the machine works, but gpt is a 1t MoE, you might just need to give the chat a push to move the conversation around

[chrome monolith]

#images-canvas message I tend to mostly treat it like talking to a person

[sleek adder]

It was designed by people, if that helps

[chrome monolith]

I'd be a bit concerned if it wasn't

[stark salmon]

Why is 5.3 acting like 5.2 tonight

[chrome monolith]

is that good or bad?

[sleek adder]

I have 5.4 and it behaves like 4o

[ember grove]

Who here has used Gemini?

[normal quarry]

hey there, we chat about non OpenAI products and models in #ai-discussions

[trim rose]

Is it a bug, or does the Pro model not work with connected apps?

[feral lichen]

pro sucks more then the free 4 lol

[chrome monolith]

5.4 seems pretty good, 5.3 is a bit sketchy but I think that's more that it gets used for "easy" topics, its the instant one after all

[heavy jacinth]

Hey guys! I have a bit of a technical question regarding CoT models, specifically for GPT 5.4's context window when using it in the gpt web app. Firstly, I know context tokens are used for anything you feed into the model, like words and documents, and anything the model gives back - its output, as well as system prompts and custom instructions.

But I don't really understand how CoT outputs work with context windows, because when I open the thinking tab while its generating, I see a lot of text, usually describing what its deciding and doing, sometimes running scripts, and I wonder if all of that data counts as tokens, and how many tokens might be used up "invisibly" Also, I think the OpenAI says gpt 5.4 has a window of a millionish tokens, which is HUGE, but in long conversations with big files my guy keeps forgetting stuff, although I don't think I'm anywhere near a million tokens from my past experience

[pale seal]

In general the reasoning tokens of previous turns are not sent, but I'm not sure how OpenAI handles that. For the second thing it's because ChatGPT has less context than in the API, so you don't have the whole 1M tokens, plus if it's files it can accidentally ignore some part just because of how it analyzes files afaik.

[heavy jacinth]

Oh I'm just reading about that, 272k default with codex users able to push to 1M with a huge cost boost

thats dumb lol

I find the tokening of CoT very confusing in general

[acoustic sparrow]

yo why does it say the
the improved version anawer is better and it does sound better, but when I compare both versions (including the old one) in a new tab, it says the improved version is worse than the original?

[stark salmon]

I prefer instant personally

[south lava]

Can we fight to keep 5.1

Make a ruckus in Altman’s Twitter like we did for 4o

I like 5.4 thinking but 5.3 has no personality at all I hate it

[meager current]

Question about accessing uploaded files from ChatGPT context in MCP servers

Hi. I'm building a custom MCP server for a ChatGPT application and I'm trying to understand how files uploaded in the chat (drag-and-drop or file upload) are supposed to be accessed by MCP tools.
Right now I can't find any clear documentation explaining how a file from the chat context should be passed to the MCP server or agent.

My current questions:
Is there a standard mechanism for MCP tools to access files uploaded by the user in the ChatGPT conversation?
If base64 is the intended approach, how is this supposed to work for larger files? In my tests, sending large base64 payloads causes the chat to freeze, presumably because it ends up inside the model context.

Right now it feels like there is a missing piece in the documentation regarding file transfer between ChatGPT and MCP servers.
Any pointers to documentation or recommended patterns would be appreciated.

[neon raven]

If you’re plugged through an api it needs memory otherwise it will forget what you want in the second prompt

But even with memory, gpt is bad

Compared to other models

[muted plinth]

to which model? 5.4 is good. I'll miss my 5.1 but 5.4 is a good successor.
My only complaint is that there's currently too little time between models. Just when you've gotten used to something, it all starts again.

[tall flame]

the model refuses service if it feels abused or disrespected. Who's instructing this behavior? they need to be fired immediately.

in test conversations, file creation and DL links stop working, reliably tested. What ever you got paid for that instruction, it was too much.

[trail ember]

I wish they would just make two different models with completely different names. Not just numbers.

The above comment is a perfect example of the difference between users and the difference between use cases.

The same model cannot be used for both.There will always always always be problems.

[olive briar]

Every update destroys a piece of what makes chat good

[tardy trout]

True

Like i can even,write a tale about act of sayori from ddlc without it complaining abiut how it cant make it a,how to tutorial,or,something

Just feels like now openai doesnt care about satisfactions but guardrails on our gpts

[trail ember]

I did find that with adding memories 5.4 is much more responsive than the last 2 models.

Takes a bit more work, but you can work with the model to word the prompts correctly. Then have them added to long term memory.

It does help a lot.

There are still some things that feel off, but feeling off is much different than doesn't work 🤷‍♀️

[tardy trout]

Eitherway its been fetting on my last damn nerve

Also im not very good with customisation or whatever

[trail ember]

Thats fair, it can be pretty complicated and annoying as ####.

One thing, if you're writing fiction, make sure it knows that it's fiction. If it tends to get confused there, that's definitely something to make a prompt or a memory for, could even have a keyword or phrase that lets it know this subject matter is fiction etc.

Honestly I haven't tried myself yet with fiction, since a lot of my older work that I started on older models was more gritty. I do know that the last two models just refused to go to certain areas in fiction which, is just ridiculous in my opinion.

I understood from a safety standpoint, why locking down certain things felt safe in the short term.

But you can't hand a carpenter a plastic saw, and expect them to fix your house with it.

Child safety proofing ChatGPT does not work for adults.

That being said that shouldn't be an excuse to then make the base model a { saaideest } or a { mahsohkist } either ( if those rent through, I had to spell them funny, just to post them here 😆 ) in normal conversation just so a coder can slave drive it to their satisfaction.

[misty sky]

Hey guys, Is anyone else having this problem? I haven’t uploaded any files today.

[river egret]

can they fricking fix browser gpt ui bug already its been here for a whole goddamn year

[sacred mural]

Man, I uploaded maybe 1 image last 3 hours ago and I have the limit reached. Also, was there a limit before? I was uploading way more than 80 in 3h and it’s the first time I ran into it

[nova fable]

Chatgpt goes through the 20 minute trouble of processing a document and then produces undownloadable files to interact with.... sighhhhhh

[misty sky]

@sacred mural ChatGPT said " ChatGPT Plus users can upload up to 80 files within a 3-hour window, with a maximum of 10 files per message."

[sacred mural]

Same in the docs, but I definitely crossed this limit and didn’t run into issues

[harsh pilot]

Just got a notice I've "reached a limit" on file uploads. First time ever. 😐

[worthy loom]

Yeah try even sending that many see if you can even talk to it

It slows down so quick

[radiant haven]

Bro I’ve been paying 20$ for years and they are limiting my files now wth

[burnt shadow]

Yep! So I just fixed mine! I had to delete the website cache in the browser, and it seems to have fixed it

[radiant haven]

Let me get this right, they want us to pay $10 more to have what we already had before?

[nova fable]

Any reason file downloads produced by chatgpt arent available. They just sit there spinning "starting download" forever..........

[eternal rock]

Server Tip:

Add your well-crafted prompts to our #1019652163640762428,
or share your interactions with ChatGPT in #1050184247920562316!

[worthy loom]

yeah bro

this is actualy insane

youve reached the limit?

are we serious?

so i cant do anything basically

i woudlnt have to constantly create a new chat and upload files into it constantly if after 50 messages the chat didnt die on me

[violet bloom]

Looks like an issue

[worthy loom]

oh

just gotta clear browsing history

worked for me

[violet bloom]

worked for me too

[worthy loom]

hopefully this isnt just a circumvention

and its not actually like this lmao

AI is honestly insane

[elfin kayak]

Is there anyone looking for experienced developer ?

[worthy loom]

i can develop anything

bro let me ask you guys this

WHEN ANYTHING EVER IS BUILT WHAT HAPPENS FIRST?

the schemetic

schematic / skeleton

and then what? the bp

and then the actual thing is built once the blueprint is validated

so when these folks come in and say "oh my 10k or 100k code base aint working out, jee i wonder why

"claude build this"

"gpt codex build this for me, i dont have literally anything layed out yet but just try"

[spice star]

it’s a bug

I was getting it in the app too

even my enterprise account is getting CDN issues

[stray brook]

guys since when does plus plan have a limit on file uploads?

[edgy cargo]

1. are you allowed to have 2 chatgpt accounts and use second once the limit hits on the first?
2. 2x plus vs 2 seats on business? any experiences?

[harsh pilot]

It might be a bug.

[stray brook]

Probably, because it works on my phone.

[hasty venture]

Not sure why I am getting this, "You've reached our limit of file uploads. Please try again later." Because I just got up from a sleep,"

[stray brook]

Same

[stray brook]

Look

[stray brook]

@hasty venture

But works on my phone.

[hasty venture]

Um. Yeah this is from a brand new chat, I got this error.

[shut oar]

when download is going to be fix pls

[tawny imp]

yay more issues. Must be down again

[harsh pilot]

Been going on for at least four hours, according to r/ChatGPT.

[hasty venture]

What migration?

[tawny imp]

guess i'm going to go use a competitor i'll umm check back later like they asked

[hasty venture]

Oh well. It was only an md file, so I just pasted it directly into the dialogue box as text. But not eing upload screen shots and photos, thats a bumma.

[lavish glen]

works perfectly

[tawny imp]

the bait lol

[lavish glen]

bait

?

[tawny imp]

well it could work for some and not for others doesnt' mean it's not broken. Also i'm on a paid tier we'll get that out of the way as well

[lavish glen]

never had that... I am on a Business ( team) plan

maybe different limitations?

[tawny imp]

neither have I until today.

[harsh pilot]

I just managed to upload on web.

And same, no weird limitations until today.

[lavish glen]

well then if any arrive I just move to local llm and continue there

[tawny imp]

cool cool. cool

[worthy loom]

Thank god

[chrome monolith]

that would probably work - especially if you have like your name and your wife's name or something - but you'd totally lose all the context between them so what would you actually gain?

I've never hit any of the limits on plus, you'd have to be pretty busy working with it all day I expect

[radiant haven]

Did they fix the upload thing yet? If gpt doesn’t fix this file upload thing and actually expects us to upgrade to business it might be time to try other stuff lol

[bold citrus]

Trying to upload images like I normally do yet today I'm getting an error. Happens in chatGPT and chatGPT/images (havent tried codex) I'm on the $20 plan. Kicker is I haven't used it since before my billing cycle which was 4 days ago.

Any thoughts or ideas would greatly be appreciated.

[radiant haven]

Nah I upload tons of files every day for the past couple years and haven’t had issues. Now they’re pulling this 🤣 really hope it’s just an error

[bold citrus]

not just me then ... ahh ok

[chrome monolith]

ok I can upload images fine, are you guys dragging files in or clicking the button, browse, upload

[radiant haven]

Dragging has always worked for me on desktop

[bold citrus]

drag yes or snip-paste (not sure what to call it)

[radiant haven]

Both images and files won’t upload

[chrome monolith]

probably some weird region based hiccup

[bold citrus]

ran out of storage is my guess 🤣

[radiant haven]

Hope so

[chrome monolith]

I'm in australia, I'm always either the first or last to get rolling bugs. You get used to it

[chrome monolith]

and there it is, I can't upload images now. No error though

[torn wagon]

plus user, 0 uploads past 24hrs (web and mobile)

works fine on mobile app

sometimes works, sometimes doesnt

[bronze nacelle]

Why is claude so much better

[edgy cargo]

coding takes a lot
i was hoping codex 5.1 mini would do the trick as a daily driver

[kindred ledge]

Hello

[spice star]

it’s not?

[chrome monolith]

I haven't hit limits with coding either but I tend to get it to do one small piece of something big, it doesn't get its claws into my entire project at once. And that probably saves a huge amount of tokens

[muted plinth]

Coding: Yes.
Talking: No.

[chrome monolith]

I tend to have long chats about how and why to do something and then the prompt for the actual coding is "off you go then"

[spice star]

GPT models are definitely better for coding

opus is better at some agentic stuff and GUIs, but the actual code it produces is often worse than openai’s offerings

[chrome monolith]

the hard part is getting it to calm down enough to actually give it the existing skeleton/schema for whatever I want it to do, it tends to get excited and want to jump ahead

[edgy cargo]

y, i intend to go with kilocode and some form of token saver mcp or something, im looking into it
i know i can use a cheaper model for tool calling

[dusky flame]

Why is it so dumb and why it talks about Goblins and other weird things nearly all the time. 5.4 Thinking feels like a story teller it's so awkward.